From 7745f208bc8724351896aed6bc27b8bf26a300a7 Mon Sep 17 00:00:00 2001
From: Jared Van Bortel <jared@nomic.ai>
Date: Tue, 11 Mar 2025 13:33:06 -0400
Subject: [PATCH] WIP (clang is crashing)

---
 .codespellrc                                  |   2 +-
 .gitmodules                                   |   6 +
 docs/gpt4all_desktop/settings.md              |   1 -
 gpt4all-backend/deps/CMakeLists.txt           |   2 +
 gpt4all-backend/deps/date                     |   1 +
 .../include/gpt4all-backend/formatters.h      |   9 +-
 .../include/gpt4all-backend/ollama-client.h   |  20 +-
 gpt4all-backend/src/CMakeLists.txt            |   1 +
 gpt4all-backend/src/ollama-client.cpp         |  14 +-
 gpt4all-backend/src/ollama-types.cpp          |   3 +-
 gpt4all-chat/CMakeLists.txt                   |  18 +-
 gpt4all-chat/deps/CMakeLists.txt              |   4 +
 gpt4all-chat/deps/generator                   |   1 +
 gpt4all-chat/qml/ApplicationSettings.qml      |  32 --
 gpt4all-chat/qml/ModelSettings.qml            |  47 --
 gpt4all-chat/src/chatlistmodel.cpp            |   2 +-
 gpt4all-chat/src/chatllm.cpp                  | 431 ++++++------------
 gpt4all-chat/src/chatllm.h                    |  23 +-
 gpt4all-chat/src/embllm.cpp                   |   4 -
 gpt4all-chat/src/json-helpers.cpp             |  46 ++
 gpt4all-chat/src/json-helpers.h               |  15 +
 gpt4all-chat/src/llmodel/chat.h               |  32 --
 gpt4all-chat/src/llmodel/openai.h             |  75 ---
 gpt4all-chat/src/llmodel/provider.cpp         |  26 --
 gpt4all-chat/src/llmodel/provider.h           |  47 --
 gpt4all-chat/src/llmodel_chat.h               |  34 ++
 gpt4all-chat/src/llmodel_description.cpp      |  22 +
 gpt4all-chat/src/llmodel_description.h        |  40 ++
 gpt4all-chat/src/llmodel_ollama.cpp           |  82 ++++
 gpt4all-chat/src/llmodel_ollama.h             | 122 +++++
 .../openai.cpp => llmodel_openai.cpp}         | 113 +++--
 gpt4all-chat/src/llmodel_openai.h             | 139 ++++++
 gpt4all-chat/src/llmodel_provider.cpp         | 139 ++++++
 gpt4all-chat/src/llmodel_provider.h           | 199 ++++++++
 gpt4all-chat/src/llmodel_provider.inl         |  30 ++
 gpt4all-chat/src/modellist.h                  |  11 +
 gpt4all-chat/src/mysettings.cpp               |  25 -
 gpt4all-chat/src/mysettings.h                 |  12 +-
 gpt4all-chat/src/server.cpp                   |  75 +--
 gpt4all-chat/src/store_base.cpp               | 164 +++++++
 gpt4all-chat/src/store_base.h                 | 119 +++++
 gpt4all-chat/src/store_base.inl               | 128 ++++++
 gpt4all-chat/src/store_provider.cpp           |  25 +
 gpt4all-chat/src/store_provider.h             |  49 ++
 gpt4all-chat/src/utils.h                      |  40 ++
 gpt4all-chat/src/utils.inl                    |  34 ++
 46 files changed, 1760 insertions(+), 704 deletions(-)
 create mode 160000 gpt4all-backend/deps/date
 create mode 160000 gpt4all-chat/deps/generator
 create mode 100644 gpt4all-chat/src/json-helpers.cpp
 create mode 100644 gpt4all-chat/src/json-helpers.h
 delete mode 100644 gpt4all-chat/src/llmodel/chat.h
 delete mode 100644 gpt4all-chat/src/llmodel/openai.h
 delete mode 100644 gpt4all-chat/src/llmodel/provider.cpp
 delete mode 100644 gpt4all-chat/src/llmodel/provider.h
 create mode 100644 gpt4all-chat/src/llmodel_chat.h
 create mode 100644 gpt4all-chat/src/llmodel_description.cpp
 create mode 100644 gpt4all-chat/src/llmodel_description.h
 create mode 100644 gpt4all-chat/src/llmodel_ollama.cpp
 create mode 100644 gpt4all-chat/src/llmodel_ollama.h
 rename gpt4all-chat/src/{llmodel/openai.cpp => llmodel_openai.cpp} (59%)
 create mode 100644 gpt4all-chat/src/llmodel_openai.h
 create mode 100644 gpt4all-chat/src/llmodel_provider.cpp
 create mode 100644 gpt4all-chat/src/llmodel_provider.h
 create mode 100644 gpt4all-chat/src/llmodel_provider.inl
 create mode 100644 gpt4all-chat/src/store_base.cpp
 create mode 100644 gpt4all-chat/src/store_base.h
 create mode 100644 gpt4all-chat/src/store_base.inl
 create mode 100644 gpt4all-chat/src/store_provider.cpp
 create mode 100644 gpt4all-chat/src/store_provider.h

diff --git a/.codespellrc b/.codespellrc
index 0f401f6b..9a625e32 100644
--- a/.codespellrc
+++ b/.codespellrc
@@ -1,3 +1,3 @@
 [codespell]
-ignore-words-list = blong, afterall, assistent, crasher, requestor
+ignore-words-list = blong, afterall, assistent, crasher, requestor, nam
 skip = ./.git,./gpt4all-chat/translations,*.pdf,*.svg,*.lock
diff --git a/.gitmodules b/.gitmodules
index 9549e256..6afea7f0 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -26,3 +26,9 @@
 [submodule "gpt4all-backend/deps/qcoro"]
 	path = deps/qcoro
 	url = https://github.com/nomic-ai/qcoro.git
+[submodule "gpt4all-backend/deps/date"]
+	path = gpt4all-backend/deps/date
+	url = https://github.com/HowardHinnant/date.git
+[submodule "gpt4all-chat/deps/generator"]
+	path = gpt4all-chat/deps/generator
+	url = https://github.com/TartanLlama/generator.git
diff --git a/docs/gpt4all_desktop/settings.md b/docs/gpt4all_desktop/settings.md
index e9d5eb85..d48fe069 100644
--- a/docs/gpt4all_desktop/settings.md
+++ b/docs/gpt4all_desktop/settings.md
@@ -49,7 +49,6 @@ You can **clone** an existing model, which allows you to save a configuration of
     |----------------------------|------------------------------------------|-----------|
     | **Context Length**         | Maximum length of input sequence in tokens        | 2048      |
     | **Max Length**             | Maximum length of response in tokens     | 4096      |
-    | **Prompt Batch Size**      | Token batch size for parallel processing | 128      |
     | **Temperature**            | Lower temperature gives more likely generations | 0.7       |
     | **Top P**                  | Prevents choosing highly unlikely tokens  | 0.4       |
     | **Top K**                  | Size of selection pool for tokens         | 40        |
diff --git a/gpt4all-backend/deps/CMakeLists.txt b/gpt4all-backend/deps/CMakeLists.txt
index 382e83e3..e74dcbe7 100644
--- a/gpt4all-backend/deps/CMakeLists.txt
+++ b/gpt4all-backend/deps/CMakeLists.txt
@@ -12,3 +12,5 @@ FetchContent_Declare(
     URL_HASH "SHA256=7da75f171837577a52bbf217e17f8ea576c7c246e4594d617bfde7fafd408be5"
 )
 FetchContent_MakeAvailable(boost)
+
+add_subdirectory(date)
diff --git a/gpt4all-backend/deps/date b/gpt4all-backend/deps/date
new file mode 160000
index 00000000..5bdb7e6f
--- /dev/null
+++ b/gpt4all-backend/deps/date
@@ -0,0 +1 @@
+Subproject commit 5bdb7e6f31fac909c090a46dbd9fea27b6e609a4
diff --git a/gpt4all-backend/include/gpt4all-backend/formatters.h b/gpt4all-backend/include/gpt4all-backend/formatters.h
index e64f55b4..84b8c5ba 100644
--- a/gpt4all-backend/include/gpt4all-backend/formatters.h
+++ b/gpt4all-backend/include/gpt4all-backend/formatters.h
@@ -26,7 +26,8 @@
         }                                                                        \
     }
 
-MAKE_FORMATTER(QUtf8StringView, value                    );
-MAKE_FORMATTER(QStringView,     value.toUtf8()           );
-MAKE_FORMATTER(QString,         value.toUtf8()           );
-MAKE_FORMATTER(QVariant,        value.toString().toUtf8());
+MAKE_FORMATTER(QLatin1StringView, value                    );
+MAKE_FORMATTER(QString,           value.toUtf8()           );
+MAKE_FORMATTER(QStringView,       value.toUtf8()           );
+MAKE_FORMATTER(QUtf8StringView,   value                    );
+MAKE_FORMATTER(QVariant,          value.toString().toUtf8());
diff --git a/gpt4all-backend/include/gpt4all-backend/ollama-client.h b/gpt4all-backend/include/gpt4all-backend/ollama-client.h
index 4c251024..c891b64f 100644
--- a/gpt4all-backend/include/gpt4all-backend/ollama-client.h
+++ b/gpt4all-backend/include/gpt4all-backend/ollama-client.h
@@ -11,7 +11,6 @@
 #include <QString>
 #include <QUrl>
 
-#include <cassert>
 #include <expected>
 #include <utility>
 #include <variant>
@@ -26,26 +25,21 @@ namespace gpt4all::backend {
 struct ResponseError {
 public:
     struct BadStatus { int code; };
-
-private:
     using ErrorCode = std::variant<
         QNetworkReply::NetworkError,
         boost::system::error_code,
         BadStatus
     >;
 
-public:
-    ErrorCode error;
-    QString   errorString;
-
     ResponseError(const QRestReply *reply);
+    ResponseError(const boost::system::system_error &e);
 
-    ResponseError(const boost::system::system_error &e)
-        : error(e.code())
-        , errorString(QString::fromUtf8(e.what()))
-    {
-        assert(e.code());
-    }
+    const ErrorCode &error      () { return m_error;       }
+    const QString   &errorString() { return m_errorString; }
+
+private:
+    ErrorCode m_error;
+    QString   m_errorString;
 };
 
 template <typename T>
diff --git a/gpt4all-backend/src/CMakeLists.txt b/gpt4all-backend/src/CMakeLists.txt
index 212264c3..5fc030ad 100644
--- a/gpt4all-backend/src/CMakeLists.txt
+++ b/gpt4all-backend/src/CMakeLists.txt
@@ -21,6 +21,7 @@ target_link_libraries(${TARGET} PUBLIC
 )
 target_link_libraries(${TARGET} PRIVATE
     QCoro6::Network
+    date::date
     fmt::fmt
 )
 
diff --git a/gpt4all-backend/src/ollama-client.cpp b/gpt4all-backend/src/ollama-client.cpp
index eda623d8..d32d2ae6 100644
--- a/gpt4all-backend/src/ollama-client.cpp
+++ b/gpt4all-backend/src/ollama-client.cpp
@@ -13,7 +13,6 @@
 #include <QVariant>
 #include <QtAssert>
 
-#include <coroutine>
 #include <expected>
 #include <memory>
 
@@ -28,13 +27,20 @@ namespace gpt4all::backend {
 ResponseError::ResponseError(const QRestReply *reply)
 {
     if (reply->hasError()) {
-        error = reply->networkReply()->error();
+        m_error = reply->networkReply()->error();
     } else if (!reply->isHttpStatusSuccess()) {
-        error = BadStatus(reply->httpStatus());
+        m_error = BadStatus(reply->httpStatus());
     } else
         Q_UNREACHABLE();
 
-    errorString = restErrorString(*reply);
+    m_errorString = restErrorString(*reply);
+}
+
+ResponseError::ResponseError(const boost::system::system_error &e)
+    : m_error(e.code())
+    , m_errorString(QString::fromUtf8(e.what()))
+{
+    Q_ASSERT(e.code());
 }
 
 QNetworkRequest OllamaClient::makeRequest(const QString &path) const
diff --git a/gpt4all-backend/src/ollama-types.cpp b/gpt4all-backend/src/ollama-types.cpp
index d4f9addb..c3a0eaf5 100644
--- a/gpt4all-backend/src/ollama-types.cpp
+++ b/gpt4all-backend/src/ollama-types.cpp
@@ -4,6 +4,7 @@
 
 #include <fmt/chrono.h> // IWYU pragma: keep
 #include <fmt/format.h>
+#include <date/date.h>
 
 #include <sstream>
 #include <string>
@@ -40,7 +41,7 @@ Time tag_invoke(const json::value_to_tag<Time> &, const json::value &value)
 
     Time time;
     std::istringstream iss(json::string_view(value.as_string()));
-    iss >> std::chrono::parse("%FT%T%Ez", time);
+    iss >> date::parse("%FT%T%Ez", time);
     if (!iss && !iss.eof())
         throw sys::system_error(std::make_error_code(std::errc::invalid_argument), __func__);
     return time;
diff --git a/gpt4all-chat/CMakeLists.txt b/gpt4all-chat/CMakeLists.txt
index 07485bf7..6783efc0 100644
--- a/gpt4all-chat/CMakeLists.txt
+++ b/gpt4all-chat/CMakeLists.txt
@@ -227,9 +227,6 @@ if (APPLE)
 endif()
 
 qt_add_executable(chat
-    src/llmodel/provider.cpp      src/llmodel/provider.h
-    src/llmodel/openai.cpp        src/llmodel/openai.h
-    src/main.cpp
     src/chat.cpp                  src/chat.h
     src/chatlistmodel.cpp         src/chatlistmodel.h
     src/chatllm.cpp               src/chatllm.h
@@ -241,14 +238,22 @@ qt_add_executable(chat
     src/embllm.cpp                src/embllm.h
     src/jinja_helpers.cpp         src/jinja_helpers.h
     src/jinja_replacements.cpp    src/jinja_replacements.h
+    src/json-helpers.cpp          src/json-helpers.h
     src/llm.cpp                   src/llm.h
+    src/llmodel_chat.h
+    src/llmodel_ollama.cpp        src/llmodel_ollama.h
+    src/llmodel_openai.cpp        src/llmodel_openai.h
+    src/llmodel_provider.cpp      src/llmodel_provider.h
     src/localdocs.cpp             src/localdocs.h
     src/localdocsmodel.cpp        src/localdocsmodel.h
     src/logger.cpp                src/logger.h
+    src/main.cpp
     src/modellist.cpp             src/modellist.h
     src/mysettings.cpp            src/mysettings.h
     src/network.cpp               src/network.h
     src/server.cpp                src/server.h
+    src/store_base.cpp            src/store_base.h
+    src/store_provider.cpp        src/store_provider.h
     src/tool.cpp                  src/tool.h
     src/toolcallparser.cpp        src/toolcallparser.h
     src/toolmodel.cpp             src/toolmodel.h
@@ -448,8 +453,9 @@ target_compile_definitions(chat PRIVATE QT_NO_SIGNALS_SLOTS_KEYWORDS)
 target_include_directories(chat PRIVATE deps/usearch/include
                                         deps/usearch/fp16/include)
 
-target_link_libraries(chat
-    PRIVATE Qt6::Core Qt6::HttpServer Qt6::Quick Qt6::Sql Qt6::Svg)
+target_link_libraries(chat PRIVATE
+    Qt6::Core Qt6::HttpServer Qt6::Quick Qt6::Sql Qt6::Svg
+)
 if (GPT4ALL_USING_QTPDF)
     target_compile_definitions(chat PRIVATE GPT4ALL_USE_QTPDF)
     target_link_libraries(chat PRIVATE Qt6::Pdf)
@@ -458,6 +464,7 @@ else()
     target_link_libraries(chat PRIVATE pdfium)
 endif()
 target_link_libraries(chat PRIVATE
+    Boost::describe Boost::json Boost::system
     QCoro6::Core QCoro6::Network
     QXlsx
     SingleApplication
@@ -466,6 +473,7 @@ target_link_libraries(chat PRIVATE
     gpt4all-backend
     llmodel
     nlohmann_json::nlohmann_json
+    tl::generator
 )
 target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/minja/include)
 
diff --git a/gpt4all-chat/deps/CMakeLists.txt b/gpt4all-chat/deps/CMakeLists.txt
index c8f865fa..6cec1227 100644
--- a/gpt4all-chat/deps/CMakeLists.txt
+++ b/gpt4all-chat/deps/CMakeLists.txt
@@ -14,6 +14,10 @@ add_subdirectory(QXlsx/QXlsx)
 
 add_subdirectory(json)  # required by minja
 
+# TartanLlama
+set(FUNCTION_REF_ENABLE_TESTS OFF)
+add_subdirectory(generator)
+
 if (NOT GPT4ALL_USING_QTPDF)
     # If we do not use QtPDF, we need to get PDFium.
     set(GPT4ALL_PDFIUM_TAG "chromium/6996")
diff --git a/gpt4all-chat/deps/generator b/gpt4all-chat/deps/generator
new file mode 160000
index 00000000..2a912502
--- /dev/null
+++ b/gpt4all-chat/deps/generator
@@ -0,0 +1 @@
+Subproject commit 2a912502de4f97dcba4f95c958ee0ddf7bc22cf5
diff --git a/gpt4all-chat/qml/ApplicationSettings.qml b/gpt4all-chat/qml/ApplicationSettings.qml
index c0025615..db78a82d 100644
--- a/gpt4all-chat/qml/ApplicationSettings.qml
+++ b/gpt4all-chat/qml/ApplicationSettings.qml
@@ -454,38 +454,6 @@ MySettingsTab {
             }
         }
 
-        MySettingsLabel {
-            id: nThreadsLabel
-            text: qsTr("CPU Threads")
-            helpText: qsTr("The number of CPU threads used for inference and embedding.")
-            Layout.row: 11
-            Layout.column: 0
-        }
-        MyTextField {
-            text: MySettings.threadCount
-            color: theme.textColor
-            font.pixelSize: theme.fontSizeLarge
-            Layout.alignment: Qt.AlignRight
-            Layout.row: 11
-            Layout.column: 2
-            Layout.minimumWidth: 200
-            Layout.maximumWidth: 200
-            validator: IntValidator {
-                bottom: 1
-            }
-            onEditingFinished: {
-                var val = parseInt(text)
-                if (!isNaN(val)) {
-                    MySettings.threadCount = val
-                    focus = false
-                } else {
-                    text = MySettings.threadCount
-                }
-            }
-            Accessible.role: Accessible.EditableText
-            Accessible.name: nThreadsLabel.text
-            Accessible.description: ToolTip.text
-        }
         MySettingsLabel {
             id: trayLabel
             text: qsTr("Enable System Tray")
diff --git a/gpt4all-chat/qml/ModelSettings.qml b/gpt4all-chat/qml/ModelSettings.qml
index 62906440..5374630d 100644
--- a/gpt4all-chat/qml/ModelSettings.qml
+++ b/gpt4all-chat/qml/ModelSettings.qml
@@ -791,53 +791,6 @@ MySettingsTab {
                 Accessible.description: ToolTip.text
             }
 
-            MySettingsLabel {
-                id: batchSizeLabel
-                visible: !root.currentModelInfo.isOnline
-                text: qsTr("Prompt Batch Size")
-                helpText: qsTr("The batch size used for prompt processing.")
-                Layout.row: 1
-                Layout.column: 0
-                Layout.maximumWidth: 300 * theme.fontScale
-            }
-            MyTextField {
-                id: batchSizeField
-                visible: !root.currentModelInfo.isOnline
-                text: root.currentModelInfo.promptBatchSize
-                color: theme.textColor
-                font.pixelSize: theme.fontSizeLarge
-                ToolTip.text: qsTr("Amount of prompt tokens to process at once.\nNOTE: Higher values can speed up reading prompts but will use more RAM.")
-                ToolTip.visible: hovered
-                Layout.row: 1
-                Layout.column: 1
-                validator: IntValidator {
-                    bottom: 1
-                }
-                Connections {
-                    target: MySettings
-                    function onPromptBatchSizeChanged() {
-                        batchSizeField.text = root.currentModelInfo.promptBatchSize;
-                    }
-                }
-                Connections {
-                    target: root
-                    function onCurrentModelInfoChanged() {
-                        batchSizeField.text = root.currentModelInfo.promptBatchSize;
-                    }
-                }
-                onEditingFinished: {
-                    var val = parseInt(text)
-                    if (!isNaN(val)) {
-                        MySettings.setModelPromptBatchSize(root.currentModelInfo, val)
-                        focus = false
-                    } else {
-                        text = root.currentModelInfo.promptBatchSize
-                    }
-                }
-                Accessible.role: Accessible.EditableText
-                Accessible.name: batchSizeLabel.text
-                Accessible.description: ToolTip.text
-            }
             MySettingsLabel {
                 id: repeatPenaltyLabel
                 visible: !root.currentModelInfo.isOnline
diff --git a/gpt4all-chat/src/chatlistmodel.cpp b/gpt4all-chat/src/chatlistmodel.cpp
index eac89b83..c57577cd 100644
--- a/gpt4all-chat/src/chatlistmodel.cpp
+++ b/gpt4all-chat/src/chatlistmodel.cpp
@@ -21,7 +21,7 @@
 
 
 static constexpr quint32 CHAT_FORMAT_MAGIC   = 0xF5D553CC;
-static constexpr qint32  CHAT_FORMAT_VERSION = 12;
+static constexpr qint32  CHAT_FORMAT_VERSION = 13;
 
 class MyChatListModel: public ChatListModel { };
 Q_GLOBAL_STATIC(MyChatListModel, chatListModelInstance)
diff --git a/gpt4all-chat/src/chatllm.cpp b/gpt4all-chat/src/chatllm.cpp
index 5186bc4e..66996a89 100644
--- a/gpt4all-chat/src/chatllm.cpp
+++ b/gpt4all-chat/src/chatllm.cpp
@@ -3,8 +3,9 @@
 #include "chat.h"
 #include "chatmodel.h"
 #include "jinja_helpers.h"
-#include "llmodel/chat.h"
-#include "llmodel/openai.h"
+#include "llmodel_chat.h"
+#include "llmodel_description.h"
+#include "llmodel_provider.h"
 #include "localdocs.h"
 #include "mysettings.h"
 #include "network.h"
@@ -12,6 +13,8 @@
 #include "toolcallparser.h"
 #include "toolmodel.h"
 
+#include <QCoro/QCoroAsyncGenerator>
+#include <QCoro/QCoroTask>
 #include <fmt/format.h>
 #include <gpt4all-backend/generation-params.h>
 #include <minja/minja.hpp>
@@ -118,10 +121,15 @@ public:
     virtual bool getStopGenerating () const                                                                           = 0;
 };
 
+struct PromptModelWithToolsResult {
+    ChatResponseMetadata metadata;
+    QStringList          toolCallBuffers;
+    bool                 shouldExecuteToolCall;
+};
 static auto promptModelWithTools(
-    ChatLLModel *model, BaseResponseHandler &respHandler, const backend::GenerationParams &params,
-    const QByteArray &prompt, const QStringList &toolNames
-) -> std::pair<QStringList, bool>
+    ChatLLMInstance *model, BaseResponseHandler &respHandler, const GenerationParams &params, const QByteArray &prompt,
+    const QStringList &toolNames
+) -> QCoro::Task<PromptModelWithToolsResult>
 {
     ToolCallParser toolCallParser(toolNames);
     auto handleResponse = [&toolCallParser, &respHandler](std::string_view piece) -> bool {
@@ -159,30 +167,31 @@ static auto promptModelWithTools(
 
         return !shouldExecuteToolCall && !respHandler.getStopGenerating();
     };
-    model->prompt(std::string_view(prompt), promptCallback, handleResponse, params);
+    ChatResponseMetadata metadata;
+    auto stream = model->generate(QString::fromUtf8(prompt), params, metadata);
+    QCORO_FOREACH(auto &piece, stream) {
+        if (!handleResponse(std::string_view(piece.toUtf8())))
+            break;
+    }
 
     const bool shouldExecuteToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete
         && toolCallParser.startTag() != ToolCallConstants::ThinkStartTag;
 
-    return { toolCallParser.buffers(), shouldExecuteToolCall };
+    co_return { metadata, toolCallParser.buffers(), shouldExecuteToolCall };
 }
 
 class LLModelStore {
 public:
     static LLModelStore *globalInstance();
 
-    LLModelInfo acquireModel(); // will block until llmodel is ready
-    void releaseModel(LLModelInfo &&info); // must be called when you are done
+    auto acquireModel() -> std::unique_ptr<ChatLLMInstance>; // will block until llmodel is ready
+    void releaseModel(std::unique_ptr<ChatLLMInstance> &&info); // must be called when you are done
     void destroy();
 
 private:
-    LLModelStore()
-    {
-        // seed with empty model
-        m_availableModel = LLModelInfo();
-    }
-    ~LLModelStore() {}
-    std::optional<LLModelInfo> m_availableModel;
+    LLModelStore() { m_availableModel.emplace(); /* seed with empty model */ }
+    ~LLModelStore() = default;
+    std::optional<std::unique_ptr<ChatLLMInstance>> m_availableModel;
     QMutex m_mutex;
     QWaitCondition m_condition;
     friend class MyLLModelStore;
@@ -195,7 +204,7 @@ LLModelStore *LLModelStore::globalInstance()
     return storeInstance();
 }
 
-LLModelInfo LLModelStore::acquireModel()
+auto LLModelStore::acquireModel() -> std::unique_ptr<ChatLLMInstance>
 {
     QMutexLocker locker(&m_mutex);
     while (!m_availableModel)
@@ -205,7 +214,7 @@ LLModelInfo LLModelStore::acquireModel()
     return first;
 }
 
-void LLModelStore::releaseModel(LLModelInfo &&info)
+void LLModelStore::releaseModel(std::unique_ptr<ChatLLMInstance> &&info)
 {
     QMutexLocker locker(&m_mutex);
     Q_ASSERT(!m_availableModel);
@@ -219,11 +228,6 @@ void LLModelStore::destroy()
     m_availableModel.reset();
 }
 
-void LLModelInfo::resetModel(ChatLLM *cllm, ChatLLModel *model) {
-    this->model.reset(model);
-    emit cllm->loadedModelInfoChanged();
-}
-
 ChatLLM::ChatLLM(Chat *parent, bool isServer)
     : QObject{nullptr}
     , m_chat(parent)
@@ -264,9 +268,8 @@ void ChatLLM::destroy()
 
     // The only time we should have a model loaded here is on shutdown
     // as we explicitly unload the model in all other circumstances
-    if (isModelLoaded()) {
-        m_llModelInfo.resetModel(this);
-    }
+    if (isModelLoaded())
+        m_llmInstance.reset();
 }
 
 void ChatLLM::destroyStore()
@@ -288,7 +291,7 @@ bool ChatLLM::loadDefaultModel()
         emit modelLoadingError(u"Could not find any model to load"_s);
         return false;
     }
-    return loadModel(defaultModel);
+    return QCoro::waitFor(loadModel(defaultModel));
 }
 
 void ChatLLM::trySwitchContextOfLoadedModel(const ModelInfo &modelInfo)
@@ -305,24 +308,21 @@ void ChatLLM::trySwitchContextOfLoadedModel(const ModelInfo &modelInfo)
         return;
     }
 
-    QString filePath = modelInfo.dirpath + modelInfo.filename();
-    QFileInfo fileInfo(filePath);
-
     acquireModel();
 #if defined(DEBUG_MODEL_LOADING)
-        qDebug() << "acquired model from store" << m_llmThread.objectName() << m_llModelInfo.model.get();
+        qDebug() << "acquired model from store" << m_llmThread.objectName() << m_llmInstance.get();
 #endif
 
     // The store gave us no already loaded model, the wrong type of model, then give it back to the
     // store and fail
-    if (!m_llModelInfo.model || m_llModelInfo.fileInfo != fileInfo || !m_shouldBeLoaded) {
-        LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
+    if (!m_llmInstance || *m_llmInstance->description() != *modelInfo.modelDesc() || !m_shouldBeLoaded) {
+        LLModelStore::globalInstance()->releaseModel(std::move(m_llmInstance));
         emit trySwitchContextOfLoadedModelCompleted(0);
         return;
     }
 
 #if defined(DEBUG_MODEL_LOADING)
-    qDebug() << "store had our model" << m_llmThread.objectName() << m_llModelInfo.model.get();
+    qDebug() << "store had our model" << m_llmThread.objectName() << m_llmInstance.model.get();
 #endif
 
     emit trySwitchContextOfLoadedModelCompleted(2);
@@ -330,233 +330,119 @@ void ChatLLM::trySwitchContextOfLoadedModel(const ModelInfo &modelInfo)
     emit trySwitchContextOfLoadedModelCompleted(0);
 }
 
-bool ChatLLM::loadModel(const ModelInfo &modelInfo)
+// TODO: always call with a resource guard held since this didn't previously use coroutines
+auto ChatLLM::loadModel(const ModelInfo &modelInfo) -> QCoro::Task<bool>
 {
-    // This is a complicated method because N different possible threads are interested in the outcome
-    // of this method. Why? Because we have a main/gui thread trying to monitor the state of N different
-    // possible chat threads all vying for a single resource - the currently loaded model - as the user
-    // switches back and forth between chats. It is important for our main/gui thread to never block
-    // but simultaneously always have up2date information with regards to which chat has the model loaded
-    // and what the type and name of that model is. I've tried to comment extensively in this method
-    // to provide an overview of what we're doing here.
-
-    if (isModelLoaded() && this->modelInfo() == modelInfo) {
+    // TODO: get the description from somewhere
+    bool alreadyAcquired = isModelLoaded();
+    if (alreadyAcquired && *modelInfo.modelDesc() == *m_modelInfo.modelDesc()) {
         // already acquired -> keep it
-        return true; // already loaded
+        if (modelInfo != m_modelInfo) {
+            // switch to different clone of same model
+            Q_ASSERT(modelInfo.isClone() || m_modelInfo.isClone());
+            m_modelInfo = modelInfo;
+        }
+        co_return true;
     }
 
     // reset status
     emit modelLoadingPercentageChanged(std::numeric_limits<float>::min()); // small non-zero positive value
     emit modelLoadingError("");
 
-    QString filePath = modelInfo.dirpath + modelInfo.filename();
-    QFileInfo fileInfo(filePath);
-
-    // We have a live model, but it isn't the one we want
-    bool alreadyAcquired = isModelLoaded();
     if (alreadyAcquired) {
-#if defined(DEBUG_MODEL_LOADING)
-        qDebug() << "already acquired model deleted" << m_llmThread.objectName() << m_llModelInfo.model.get();
-#endif
-        m_llModelInfo.resetModel(this);
-    } else if (!m_isServer) {
-        // This is a blocking call that tries to retrieve the model we need from the model store.
-        // If it succeeds, then we just have to restore state. If the store has never had a model
-        // returned to it, then the modelInfo.model pointer should be null which will happen on startup
-        acquireModel();
-#if defined(DEBUG_MODEL_LOADING)
-        qDebug() << "acquired model from store" << m_llmThread.objectName() << m_llModelInfo.model.get();
-#endif
-        // At this point it is possible that while we were blocked waiting to acquire the model from the
-        // store, that our state was changed to not be loaded. If this is the case, release the model
-        // back into the store and quit loading
+        // we own a different model -> destroy it and load the requested one
+        m_llmInstance.reset();
+    } else if (!m_isServer) { // (the server loads models lazily rather than eagerly)
+        // wait for the model to become available
+        acquireModel(); // (blocks)
+
+        // check if request was canceled while we were waiting
         if (!m_shouldBeLoaded) {
-#if defined(DEBUG_MODEL_LOADING)
-            qDebug() << "no longer need model" << m_llmThread.objectName() << m_llModelInfo.model.get();
-#endif
-            LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
+            LLModelStore::globalInstance()->releaseModel(std::move(m_llmInstance));
             emit modelLoadingPercentageChanged(0.0f);
-            return false;
+            co_return false;
         }
 
-        // Check if the store just gave us exactly the model we were looking for
-        if (m_llModelInfo.model && m_llModelInfo.fileInfo == fileInfo) {
-#if defined(DEBUG_MODEL_LOADING)
-            qDebug() << "store had our model" << m_llmThread.objectName() << m_llModelInfo.model.get();
-#endif
+        // if it was the requested model, we are done
+        if (m_llmInstance && *m_llmInstance->description() == *modelInfo.modelDesc()) {
             emit modelLoadingPercentageChanged(1.0f);
             setModelInfo(modelInfo);
             Q_ASSERT(!m_modelInfo.filename().isEmpty());
             if (m_modelInfo.filename().isEmpty())
                 emit modelLoadingError(u"Modelinfo is left null for %1"_s.arg(modelInfo.filename()));
-            return true;
-        } else {
-            // Release the memory since we have to switch to a different model.
-#if defined(DEBUG_MODEL_LOADING)
-            qDebug() << "deleting model" << m_llmThread.objectName() << m_llModelInfo.model.get();
-#endif
-            m_llModelInfo.resetModel(this);
+            co_return true;
         }
+
+        // we own a different model -> destroy it and load the requested one
+        m_llmInstance.reset();
     }
 
-    // Guarantee we've released the previous models memory
-    Q_ASSERT(!m_llModelInfo.model);
+    QVariantMap modelLoadProps;
+    if (!co_await loadNewModel(modelInfo, modelLoadProps))
+        co_return false; // m_shouldBeLoaded became false
 
-    // Store the file info in the modelInfo in case we have an error loading
-    m_llModelInfo.fileInfo = fileInfo;
+    emit modelLoadingPercentageChanged(isModelLoaded() ? 1.0f : 0.0f);
+    emit loadedModelInfoChanged();
 
-    if (fileInfo.exists()) {
-        QVariantMap modelLoadProps;
-        if (modelInfo.isOnline) {
-            QString apiKey;
-            QString requestUrl;
-            QString modelName;
-            {
-                QFile file(filePath);
-                bool success = file.open(QIODeviceBase::ReadOnly);
-                (void)success;
-                Q_ASSERT(success);
-                QJsonDocument doc = QJsonDocument::fromJson(file.readAll());
-                QJsonObject obj = doc.object();
-                apiKey = obj["apiKey"].toString();
-                modelName = obj["modelName"].toString();
-                if (modelInfo.isCompatibleApi) {
-                    QString baseUrl(obj["baseUrl"].toString());
-                    QUrl apiUrl(QUrl::fromUserInput(baseUrl));
-                    if (!Network::isHttpUrlValid(apiUrl)) {
-                        return false;
-                    }
-                    QString currentPath(apiUrl.path());
-                    QString suffixPath("%1/chat/completions");
-                    apiUrl.setPath(suffixPath.arg(currentPath));
-                    requestUrl = apiUrl.toString();
-                } else {
-                    requestUrl = modelInfo.url();
-                }
-            }
-            m_llModelType = LLModelTypeV1::API;
-            ChatAPI *model = new ChatAPI();
-            model->setModelName(modelName);
-            model->setRequestURL(requestUrl);
-            model->setAPIKey(apiKey);
-            m_llModelInfo.resetModel(this, model);
-        } else if (!loadNewModel(modelInfo, modelLoadProps)) {
-            return false; // m_shouldBeLoaded became false
-        }
-#if defined(DEBUG_MODEL_LOADING)
-        qDebug() << "new model" << m_llmThread.objectName() << m_llModelInfo.model.get();
-#endif
-#if defined(DEBUG)
-        qDebug() << "modelLoadedChanged" << m_llmThread.objectName();
-        fflush(stdout);
-#endif
-        emit modelLoadingPercentageChanged(isModelLoaded() ? 1.0f : 0.0f);
-        emit loadedModelInfoChanged();
+    modelLoadProps.insert("model", modelInfo.filename());
+    Network::globalInstance()->trackChatEvent("model_load", modelLoadProps);
 
-        modelLoadProps.insert("model", modelInfo.filename());
-        Network::globalInstance()->trackChatEvent("model_load", modelLoadProps);
-    } else {
-        if (!m_isServer)
-            LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo)); // release back into the store
-        resetModel();
-        emit modelLoadingError(u"Could not find file for model %1"_s.arg(modelInfo.filename()));
-    }
-
-    if (m_llModelInfo.model)
+    if (m_llmInstance)
         setModelInfo(modelInfo);
-    return bool(m_llModelInfo.model);
+    co_return bool(m_llmInstance);
 }
 
 /* Returns false if the model should no longer be loaded (!m_shouldBeLoaded).
  * Otherwise returns true, even on error. */
-bool ChatLLM::loadNewModel(const ModelInfo &modelInfo, QVariantMap &modelLoadProps)
+auto ChatLLM::loadNewModel(const ModelInfo &modelInfo, QVariantMap &modelLoadProps) -> QCoro::Task<bool>
 {
+    auto *mysettings = MySettings::globalInstance();
+
     QElapsedTimer modelLoadTimer;
     modelLoadTimer.start();
 
-    int n_ctx = MySettings::globalInstance()->modelContextLength(modelInfo);
-    int ngl = MySettings::globalInstance()->modelGpuLayers(modelInfo);
+    // TODO: pass these as generation params
+    int n_ctx = mysettings->modelContextLength(modelInfo);
+    int ngl   = mysettings->modelGpuLayers    (modelInfo);
 
-    std::string backend = "auto";
-    QString filePath = modelInfo.dirpath + modelInfo.filename();
+    m_llmInstance = modelInfo.modelDesc()->newInstance(&m_nam);
 
-    auto construct = [this, &filePath, &modelInfo, &modelLoadProps, n_ctx]() {
-        QString constructError;
-        m_llModelInfo.resetModel(this);
-        auto *model = LLModel::Implementation::construct(filePath.toStdString(), "", n_ctx);
-        m_llModelInfo.resetModel(this, model);
-
-        if (!m_llModelInfo.model) {
-            if (!m_isServer)
-                LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
-            resetModel();
-            emit modelLoadingError(u"Error loading %1: %2"_s.arg(modelInfo.filename(), constructError));
-            return false;
-        }
-
-        m_llModelInfo.model->setProgressCallback([this](float progress) -> bool {
-            progress = std::max(progress, std::numeric_limits<float>::min()); // keep progress above zero
-            emit modelLoadingPercentageChanged(progress);
-            return m_shouldBeLoaded;
-        });
-        return true;
-    };
-
-    if (!construct())
-        return true;
-
-    if (m_llModelInfo.model->isModelBlacklisted(filePath.toStdString())) {
-        static QSet<QString> warned;
-        auto fname = modelInfo.filename();
-        if (!warned.contains(fname)) {
-            emit modelLoadingWarning(
-                u"%1 is known to be broken. Please get a replacement via the download dialog."_s.arg(fname)
-            );
-            warned.insert(fname); // don't warn again until restart
-        }
-    }
-
-    bool success = m_llModelInfo.model->loadModel(filePath.toStdString(), n_ctx, ngl);
+    // TODO: progress callback
+#if 0
+    m_llmInstance->setProgressCallback([this](float progress) -> bool {
+        progress = std::max(progress, std::numeric_limits<float>::min()); // keep progress above zero
+        emit modelLoadingPercentageChanged(progress);
+        return m_shouldBeLoaded;
+    });
+#endif
+    co_await m_llmInstance->preload();
 
     if (!m_shouldBeLoaded) {
-        m_llModelInfo.resetModel(this);
+        m_llmInstance.reset();
         if (!m_isServer)
-            LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
+            LLModelStore::globalInstance()->releaseModel(std::move(m_llmInstance));
         resetModel();
         emit modelLoadingPercentageChanged(0.0f);
-        return false;
+        co_return false;
     }
 
+    bool success = true; // TODO: check for failure
     if (!success) {
-        m_llModelInfo.resetModel(this);
+        m_llmInstance.reset();
         if (!m_isServer)
-            LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
+            LLModelStore::globalInstance()->releaseModel(std::move(m_llmInstance));
         resetModel();
         emit modelLoadingError(u"Could not load model due to invalid model file for %1"_s.arg(modelInfo.filename()));
         modelLoadProps.insert("error", "loadmodel_failed");
-        return true;
-    }
-
-    switch (m_llModelInfo.model->implementation().modelType()[0]) {
-    case 'L': m_llModelType = LLModelTypeV1::LLAMA; break;
-    default:
-        {
-            m_llModelInfo.resetModel(this);
-            if (!m_isServer)
-                LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
-            resetModel();
-            emit modelLoadingError(u"Could not determine model type for %1"_s.arg(modelInfo.filename()));
-        }
+        co_return true;
     }
 
     modelLoadProps.insert("$duration", modelLoadTimer.elapsed() / 1000.);
-    return true;
+    co_return true;
 }
 
 bool ChatLLM::isModelLoaded() const
-{
-    return m_llModelInfo.model && m_llModelInfo.model->isModelLoaded();
-}
+{ return bool(m_llmInstance); }
 
 static QString &removeLeadingWhitespace(QString &s)
 {
@@ -599,50 +485,34 @@ void ChatLLM::setModelInfo(const ModelInfo &modelInfo)
 }
 
 void ChatLLM::acquireModel()
-{
-    m_llModelInfo = LLModelStore::globalInstance()->acquireModel();
-    emit loadedModelInfoChanged();
-}
+{ m_llmInstance = LLModelStore::globalInstance()->acquireModel(); }
 
 void ChatLLM::resetModel()
-{
-    m_llModelInfo = {};
-    emit loadedModelInfoChanged();
-}
+{ m_llmInstance.reset(); }
 
 void ChatLLM::modelChangeRequested(const ModelInfo &modelInfo)
 {
     // ignore attempts to switch to the same model twice
     if (!isModelLoaded() || this->modelInfo() != modelInfo) {
         m_shouldBeLoaded = true;
-        loadModel(modelInfo);
+        QCoro::waitFor(loadModel(modelInfo));
     }
 }
 
-static backend::GenerationParams genParamsFromSettings(const ModelInfo &modelInfo)
-{
-    auto *mySettings = MySettings::globalInstance();
-    return {
-        .n_predict      = mySettings->modelMaxLength          (modelInfo),
-        .top_k          = mySettings->modelTopK               (modelInfo),
-        .top_p          = float(mySettings->modelTopP         (modelInfo)),
-        .min_p          = float(mySettings->modelMinP         (modelInfo)),
-        .temp           = float(mySettings->modelTemperature  (modelInfo)),
-        .n_batch        = mySettings->modelPromptBatchSize    (modelInfo),
-        .repeat_penalty = float(mySettings->modelRepeatPenalty(modelInfo)),
-        .repeat_last_n  = mySettings->modelRepeatPenaltyTokens(modelInfo),
-    };
-}
+auto ChatLLM::modelDescription() -> const ModelDescription *
+{ return m_llmInstance->description(); }
 
 void ChatLLM::prompt(const QStringList &enabledCollections)
 {
+    auto *mySettings = MySettings::globalInstance();
+
     if (!isModelLoaded()) {
         emit responseStopped(0);
         return;
     }
 
     try {
-        promptInternalChat(enabledCollections, genParamsFromSettings(m_modelInfo));
+        promptInternalChat(enabledCollections, mySettings->modelGenParams(m_modelInfo));
     } catch (const std::exception &e) {
         // FIXME(jared): this is neither translated nor serialized
         m_chatModel->setResponseValue(u"Error: %1"_s.arg(QString::fromUtf8(e.what())));
@@ -706,7 +576,6 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
     Q_ASSERT(items.size() >= 1);
 
     auto *mySettings = MySettings::globalInstance();
-    auto &model      = m_llModelInfo.model;
 
     QString chatTemplate, systemMessage;
     auto chatTemplateSetting = mySettings->modelChatTemplate(m_modelInfo);
@@ -756,8 +625,11 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
         { "add_generation_prompt", true                },
         { "toolList",              toolList            },
     };
-    for (auto &[name, token] : model->specialTokens())
+    // TODO: implement special tokens
+#if 0
+    for (auto &[name, token] : m_llmInstance->specialTokens())
         params.emplace(std::move(name), std::move(token));
+#endif
 
     try {
         auto tmpl = loadJinjaTemplate(chatTemplate.toStdString());
@@ -769,7 +641,7 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
     Q_UNREACHABLE();
 }
 
-auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const backend::GenerationParams &params,
+auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const GenerationParams &params,
                                  qsizetype startOffset) -> ChatPromptResult
 {
     Q_ASSERT(isModelLoaded());
@@ -876,10 +748,9 @@ private:
 };
 
 auto ChatLLM::promptInternal(
-    const std::variant<std::span<const MessageItem>, std::string_view> &prompt,
-    const backend::GenerationParams params,
+    const std::variant<std::span<const MessageItem>, std::string_view> &prompt, const GenerationParams &params,
     bool usedLocalDocs
-) -> PromptResult
+) -> QCoro::Task<PromptResult>
 {
     Q_ASSERT(isModelLoaded());
 
@@ -897,22 +768,6 @@ auto ChatLLM::promptInternal(
         conversation = jinjaBuffer;
     }
 
-    // check for overlength last message
-    if (!dynamic_cast<const ChatAPI *>(m_llModelInfo.model.get())) {
-        auto nCtx = m_llModelInfo.model->contextLength();
-        std::string jinjaBuffer2;
-        auto lastMessageRendered = (messageItems && messageItems->size() > 1)
-            ? std::string_view(jinjaBuffer2 = applyJinjaTemplate({ &messageItems->back(), 1 }))
-            : conversation;
-        int32_t lastMessageLength = m_llModelInfo.model->countPromptTokens(lastMessageRendered);
-        if (auto limit = nCtx - 4; lastMessageLength > limit) {
-            throw std::invalid_argument(
-                tr("Your message was too long and could not be processed (%1 > %2). "
-                   "Please try again with something shorter.").arg(lastMessageLength).arg(limit).toUtf8().constData()
-            );
-        }
-    }
-
     PromptResult result {};
 
     QElapsedTimer totalTime;
@@ -920,16 +775,14 @@ auto ChatLLM::promptInternal(
     ChatViewResponseHandler respHandler(this, &totalTime, &result);
 
     m_timer->start();
-    QStringList finalBuffers;
-    bool        shouldExecuteTool;
+    PromptModelWithToolsResult withToolsResult;
     try {
         emit promptProcessing();
-        m_llModelInfo.model->setThreadCount(mySettings->threadCount());
         m_stopGenerating = false;
         // TODO: set result.promptTokens based on ollama prompt_eval_count
         // TODO: support interruption via m_stopGenerating
-        std::tie(finalBuffers, shouldExecuteTool) = promptModelWithTools(
-            m_llModelInfo.model.get(), handlePrompt, respHandler, params,
+        withToolsResult = co_await promptModelWithTools(
+            m_llmInstance.get(), respHandler, params,
             QByteArray::fromRawData(conversation.data(), conversation.size()),
             ToolCallConstants::AllTagNames
         );
@@ -937,6 +790,8 @@ auto ChatLLM::promptInternal(
         m_timer->stop();
         throw;
     }
+    // TODO: use metadata
+    auto &[metadata, finalBuffers, shouldExecuteTool] = withToolsResult;
 
     m_timer->stop();
     qint64 elapsed = totalTime.elapsed();
@@ -964,13 +819,13 @@ auto ChatLLM::promptInternal(
     else
         emit responseStopped(elapsed);
 
-    return result;
+    co_return result;
 }
 
 void ChatLLM::setShouldBeLoaded(bool b)
 {
 #if defined(DEBUG_MODEL_LOADING)
-    qDebug() << "setShouldBeLoaded" << m_llmThread.objectName() << b << m_llModelInfo.model.get();
+    qDebug() << "setShouldBeLoaded" << m_llmThread.objectName() << b << m_llmInstance.model.get();
 #endif
     m_shouldBeLoaded = b; // atomic
     emit shouldBeLoadedChanged();
@@ -1001,15 +856,15 @@ void ChatLLM::unloadModel()
         emit modelLoadingPercentageChanged(std::numeric_limits<float>::min()); // small non-zero positive value
 
 #if defined(DEBUG_MODEL_LOADING)
-    qDebug() << "unloadModel" << m_llmThread.objectName() << m_llModelInfo.model.get();
+    qDebug() << "unloadModel" << m_llmThread.objectName() << m_llmInstance.model.get();
 #endif
 
     if (m_forceUnloadModel) {
-        m_llModelInfo.resetModel(this);
+        m_llmInstance.reset();
         m_forceUnloadModel = false;
     }
 
-    LLModelStore::globalInstance()->releaseModel(std::move(m_llModelInfo));
+    LLModelStore::globalInstance()->releaseModel(std::move(m_llmInstance));
 }
 
 void ChatLLM::reloadModel()
@@ -1021,13 +876,13 @@ void ChatLLM::reloadModel()
         return;
 
 #if defined(DEBUG_MODEL_LOADING)
-    qDebug() << "reloadModel" << m_llmThread.objectName() << m_llModelInfo.model.get();
+    qDebug() << "reloadModel" << m_llmThread.objectName() << m_llmInstance.model.get();
 #endif
     const ModelInfo m = modelInfo();
     if (m.name().isEmpty())
         loadDefaultModel();
     else
-        loadModel(m);
+        QCoro::waitFor(loadModel(m));
 }
 
 // This class throws discards the text within thinking tags, for use with chat names and follow-up questions.
@@ -1111,8 +966,8 @@ void ChatLLM::generateName()
     try {
         // TODO: support interruption via m_stopGenerating
         promptModelWithTools(
-            m_llModelInfo.model.get(),
-            respHandler, genParamsFromSettings(m_modelInfo),
+            m_llmInstance.get(),
+            respHandler, mySettings->modelGenParams(m_modelInfo),
             applyJinjaTemplate(forkConversation(chatNamePrompt)).c_str(),
             { ToolCallConstants::ThinkTagName }
         );
@@ -1187,8 +1042,8 @@ void ChatLLM::generateQuestions(qint64 elapsed)
     try {
         // TODO: support interruption via m_stopGenerating
         promptModelWithTools(
-            m_llModelInfo.model.get(),
-            respHandler, genParamsFromSettings(m_modelInfo),
+            m_llmInstance.get(),
+            respHandler, mySettings->modelGenParams(m_modelInfo),
             applyJinjaTemplate(forkConversation(suggestedFollowUpPrompt)).c_str(),
             { ToolCallConstants::ThinkTagName }
         );
@@ -1199,39 +1054,13 @@ void ChatLLM::generateQuestions(qint64 elapsed)
     emit responseStopped(elapsed);
 }
 
-// this function serialized the cached model state to disk.
-// we want to also serialize n_ctx, and read it at load time.
 bool ChatLLM::serialize(QDataStream &stream, int version)
 {
-    if (version < 11) {
-        if (version >= 6) {
-            stream << false; // serializeKV
-        }
-        if (version >= 2) {
-            if (m_llModelType == LLModelTypeV1::NONE) {
-                qWarning() << "ChatLLM ERROR: attempted to serialize a null model for chat id" << m_chat->id()
-                           << "name" << m_chat->name();
-                return false;
-            }
-            stream << m_llModelType;
-            stream << 0; // state version
-        }
-        {
-            QString dummy;
-            stream << dummy; // response
-            stream << dummy; // generated name
-        }
-        stream << quint32(0); // prompt + response tokens
-
-        if (version < 6) { // serialize binary state
-            if (version < 4) {
-                stream << 0; // responseLogits
-            }
-            stream << int32_t(0); // n_past
-            stream << quint64(0); // input token count
-            stream << QByteArray(); // KV cache state
-        }
-    }
+    static constexpr int VERSION_MIN = 13;
+    if (version < VERSION_MIN)
+        throw std::runtime_error(fmt::format("ChatLLM does not support serializing as version {} (min is {})",
+                                             version, VERSION_MIN));
+    // nothing to do here; ChatLLM doesn't serialize any state itself anymore
     return stream.status() == QDataStream::Ok;
 }
 
diff --git a/gpt4all-chat/src/chatllm.h b/gpt4all-chat/src/chatllm.h
index 46adf231..5c2491ba 100644
--- a/gpt4all-chat/src/chatllm.h
+++ b/gpt4all-chat/src/chatllm.h
@@ -3,7 +3,7 @@
 
 #include "chatmodel.h"
 #include "database.h"
-#include "llmodel/chat.h"
+#include "llmodel_chat.h"
 #include "modellist.h"
 
 #include <QByteArray>
@@ -31,6 +31,7 @@ using namespace Qt::Literals::StringLiterals;
 
 class ChatLLM;
 class QDataStream;
+namespace QCoro { template <typename T> class Task; }
 
 
 // NOTE: values serialized to disk, do not change or reuse
@@ -89,12 +90,6 @@ inline LLModelTypeV1 parseLLModelTypeV0(int v0)
     }
 }
 
-struct LLModelInfo {
-    std::unique_ptr<gpt4all::ui::ChatLLModel> model;
-    QFileInfo fileInfo;
-    void resetModel(ChatLLM *cllm, gpt4all::ui::ChatLLModel *model = nullptr);
-};
-
 class TokenTimer : public QObject {
     Q_OBJECT
 public:
@@ -173,7 +168,7 @@ public Q_SLOTS:
     void prompt(const QStringList &enabledCollections);
     bool loadDefaultModel();
     void trySwitchContextOfLoadedModel(const ModelInfo &modelInfo);
-    bool loadModel(const ModelInfo &modelInfo);
+    auto loadModel(const ModelInfo &modelInfo) -> QCoro::Task<bool>;
     void modelChangeRequested(const ModelInfo &modelInfo);
     void unloadModel();
     void reloadModel();
@@ -215,14 +210,16 @@ protected:
         QList<ResultInfo> databaseResults;
     };
 
-    auto promptInternalChat(const QStringList &enabledCollections, const gpt4all::backend::GenerationParams &params,
+    auto modelDescription() -> const gpt4all::ui::ModelDescription *;
+
+    auto promptInternalChat(const QStringList &enabledCollections, const gpt4all::ui::GenerationParams &params,
                             qsizetype startOffset = 0) -> ChatPromptResult;
     // passing a string_view directly skips templating and uses the raw string
     auto promptInternal(const std::variant<std::span<const MessageItem>, std::string_view> &prompt,
-                        const gpt4all::backend::GenerationParams &params, bool usedLocalDocs) -> PromptResult;
+                        const gpt4all::ui::GenerationParams &params, bool usedLocalDocs) -> QCoro::Task<PromptResult>;
 
 private:
-    bool loadNewModel(const ModelInfo &modelInfo, QVariantMap &modelLoadProps);
+    auto loadNewModel(const ModelInfo &modelInfo, QVariantMap &modelLoadProps) -> QCoro::Task<bool>;
 
     std::vector<MessageItem> forkConversation(const QString &prompt) const;
 
@@ -237,11 +234,11 @@ protected:
 
 private:
     const Chat *m_chat;
-    LLModelInfo m_llModelInfo;
-    LLModelTypeV1 m_llModelType = LLModelTypeV1::NONE;
+    std::unique_ptr<gpt4all::ui::ChatLLMInstance> m_llmInstance;
     ModelInfo m_modelInfo;
     TokenTimer *m_timer;
     QThread m_llmThread;
+    QNetworkAccessManager m_nam; // TODO(jared): avoid making multiple thread pools
     std::atomic<bool> m_stopGenerating;
     std::atomic<bool> m_shouldBeLoaded;
     std::atomic<bool> m_forceUnloadModel;
diff --git a/gpt4all-chat/src/embllm.cpp b/gpt4all-chat/src/embllm.cpp
index e4e53820..11646afa 100644
--- a/gpt4all-chat/src/embllm.cpp
+++ b/gpt4all-chat/src/embllm.cpp
@@ -110,10 +110,6 @@ bool EmbeddingLLMWorker::loadModel()
         return false;
     }
 
-    // FIXME(jared): the user may want this to take effect without having to restart
-    int n_threads = MySettings::globalInstance()->threadCount();
-    m_model->setThreadCount(n_threads);
-
     return true;
 }
 
diff --git a/gpt4all-chat/src/json-helpers.cpp b/gpt4all-chat/src/json-helpers.cpp
new file mode 100644
index 00000000..2224f90d
--- /dev/null
+++ b/gpt4all-chat/src/json-helpers.cpp
@@ -0,0 +1,46 @@
+#include "json-helpers.h"
+
+#include <boost/json.hpp> // IWYU pragma: keep
+#include <boost/system.hpp> // IWYU pragma: keep
+#include <gpt4all-backend/json-helpers.h>
+
+#include <QByteArray>
+#include <QUrl>
+#include <QUuid>
+#include <QtAssert>
+
+#include <system_error>
+
+namespace json = boost::json;
+namespace sys = boost::system;
+
+
+void tag_invoke(const boost::json::value_from_tag &, boost::json::value &value, const QUuid &uuid)
+{
+    auto bytes = uuid.toRfc4122().toBase64();
+    value = json::value_from(json::string_view(bytes.data(), bytes.size()));
+}
+
+QUuid tag_invoke(const boost::json::value_to_tag<QUuid> &, const boost::json::value &value)
+{
+    auto &s = value.as_string();
+    auto bytes = QByteArray::fromRawData(s.data(), s.size());
+    auto result = QByteArray::fromBase64Encoding(bytes);
+    if (!result)
+        throw sys::system_error(std::make_error_code(std::errc::invalid_argument), __func__);
+    auto uuid = QUuid::fromRfc4122(result.decoded);
+    Q_ASSERT(!uuid.isNull()); // this may fail if the user manually creates a null UUID
+    return uuid;
+}
+
+void tag_invoke(const boost::json::value_from_tag &, boost::json::value &value, const QUrl &url)
+{
+    auto bytes = url.toEncoded();
+    value = json::value_from(json::string_view(bytes.data(), bytes.size()));
+}
+
+QUrl tag_invoke(const boost::json::value_to_tag<QUrl> &, const boost::json::value &value)
+{
+    auto &s = value.as_string();
+    return QUrl::fromEncoded(QByteArray::fromRawData(s.data(), s.size()));
+}
diff --git a/gpt4all-chat/src/json-helpers.h b/gpt4all-chat/src/json-helpers.h
new file mode 100644
index 00000000..27334b37
--- /dev/null
+++ b/gpt4all-chat/src/json-helpers.h
@@ -0,0 +1,15 @@
+#pragma once
+
+class QUrl;
+class QUuid;
+namespace boost::json {
+    class value;
+    struct value_from_tag;
+    template <typename T> struct value_to_tag;
+}
+
+void tag_invoke(const boost::json::value_from_tag &, boost::json::value &value, const QUuid &uuid);
+QUuid tag_invoke(const boost::json::value_to_tag<QUuid> &, const boost::json::value &value);
+
+void tag_invoke(const boost::json::value_from_tag &, boost::json::value &value, const QUrl &url);
+QUrl tag_invoke(const boost::json::value_to_tag<QUrl> &, const boost::json::value &value);
diff --git a/gpt4all-chat/src/llmodel/chat.h b/gpt4all-chat/src/llmodel/chat.h
deleted file mode 100644
index e13f4065..00000000
--- a/gpt4all-chat/src/llmodel/chat.h
+++ /dev/null
@@ -1,32 +0,0 @@
-#pragma once
-
-#include <QStringView>
-
-class QString;
-namespace QCoro { template <typename T> class AsyncGenerator; }
-namespace gpt4all::backend { struct GenerationParams; }
-
-
-namespace gpt4all::ui {
-
-
-struct ChatResponseMetadata {
-    int nPromptTokens;
-    int nResponseTokens;
-};
-
-// TODO: implement two of these; one based on Ollama (TBD) and the other based on OpenAI (chatapi.h)
-class ChatLLModel {
-public:
-    virtual ~ChatLLModel() = 0;
-
-    [[nodiscard]]
-    virtual QString name() = 0;
-
-    virtual void preload() = 0;
-    virtual auto chat(QStringView prompt, const backend::GenerationParams &params,
-                      /*out*/ ChatResponseMetadata &metadata) -> QCoro::AsyncGenerator<QString> = 0;
-};
-
-
-} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel/openai.h b/gpt4all-chat/src/llmodel/openai.h
deleted file mode 100644
index 3046a5c0..00000000
--- a/gpt4all-chat/src/llmodel/openai.h
+++ /dev/null
@@ -1,75 +0,0 @@
-#pragma once
-
-#include "chat.h"
-#include "provider.h"
-
-#include <QObject>
-#include <QQmlEngine>
-#include <QString>
-#include <QUrl>
-
-class QNetworkAccessManager;
-
-
-namespace gpt4all::ui {
-
-
-class OpenaiModelDescription : public QObject {
-    Q_OBJECT
-    QML_ELEMENT
-
-public:
-    explicit OpenaiModelDescription(OpenaiProvider *provider, QString displayName, QString modelName)
-        : QObject(provider)
-        , m_provider(provider)
-        , m_displayName(std::move(displayName))
-        , m_modelName(std::move(modelName))
-        {}
-
-    // getters
-    [[nodiscard]] OpenaiProvider *provider   () const { return m_provider;    }
-    [[nodiscard]] const QString  &displayName() const { return m_displayName; }
-    [[nodiscard]] const QString  &modelName  () const { return m_modelName;   }
-
-    // setters
-    void setDisplayName(QString value);
-    void setModelName  (QString value);
-
-Q_SIGNALS:
-    void displayNameChanged(const QString &value);
-    void modelNameChanged  (const QString &value);
-
-private:
-    OpenaiProvider *m_provider;
-    QString         m_displayName;
-    QString         m_modelName;
-};
-
-struct OpenaiConnectionDetails {
-    QUrl    baseUrl;
-    QString modelName;
-    QString apiKey;
-
-    OpenaiConnectionDetails(const OpenaiModelDescription *desc)
-        : baseUrl(desc->provider()->baseUrl())
-        , apiKey(desc->provider()->apiKey())
-        , modelName(desc->modelName())
-        {}
-};
-
-class OpenaiLLModel : public ChatLLModel {
-public:
-    explicit OpenaiLLModel(OpenaiConnectionDetails connDetails, QNetworkAccessManager *nam);
-
-    void preload() override { /* not supported -> no-op */ }
-
-    auto chat(QStringView prompt, const backend::GenerationParams &params, /*out*/ ChatResponseMetadata &metadata)
-        -> QCoro::AsyncGenerator<QString> override;
-
-private:
-    OpenaiConnectionDetails  m_connDetails;
-    QNetworkAccessManager   *m_nam;
-};
-
-
-} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel/provider.cpp b/gpt4all-chat/src/llmodel/provider.cpp
deleted file mode 100644
index 779925a1..00000000
--- a/gpt4all-chat/src/llmodel/provider.cpp
+++ /dev/null
@@ -1,26 +0,0 @@
-#include "provider.h"
-
-#include <utility>
-
-
-namespace gpt4all::ui {
-
-
-void OpenaiProvider::setBaseUrl(QUrl value)
-{
-    if (m_baseUrl != value) {
-        m_baseUrl = std::move(value);
-        emit baseUrlChanged(m_baseUrl);
-    }
-}
-
-void OpenaiProvider::setApiKey(QString value)
-{
-    if (m_apiKey != value) {
-        m_apiKey = std::move(value);
-        emit apiKeyChanged(m_apiKey);
-    }
-}
-
-
-} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel/provider.h b/gpt4all-chat/src/llmodel/provider.h
deleted file mode 100644
index 7189e726..00000000
--- a/gpt4all-chat/src/llmodel/provider.h
+++ /dev/null
@@ -1,47 +0,0 @@
-#pragma once
-
-#include <QObject>
-#include <QQmlEngine>
-#include <QString>
-#include <QUrl>
-
-
-namespace gpt4all::ui {
-
-
-class ModelProvider : public QObject {
-    Q_OBJECT
-
-    Q_PROPERTY(QString name READ name CONSTANT)
-
-public:
-    [[nodiscard]] virtual QString name() = 0;
-};
-
-class OpenaiProvider : public ModelProvider {
-    Q_OBJECT
-    QML_ELEMENT
-
-    Q_PROPERTY(QUrl    baseUrl READ baseUrl WRITE setBaseUrl NOTIFY baseUrlChanged)
-    Q_PROPERTY(QString apiKey  READ apiKey  WRITE setApiKey  NOTIFY apiKeyChanged)
-
-public:
-    [[nodiscard]] QString name() override  { return m_name;    }
-    [[nodiscard]] const QUrl    &baseUrl() { return m_baseUrl; }
-    [[nodiscard]] const QString &apiKey () { return m_apiKey;  }
-
-    void setBaseUrl(QUrl    value);
-    void setApiKey (QString value);
-
-Q_SIGNALS:
-    void baseUrlChanged(const QUrl    &value);
-    void apiKeyChanged (const QString &value);
-
-private:
-    QString m_name;
-    QUrl    m_baseUrl;
-    QString m_apiKey;
-};
-
-
-} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_chat.h b/gpt4all-chat/src/llmodel_chat.h
new file mode 100644
index 00000000..ca533a31
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_chat.h
@@ -0,0 +1,34 @@
+#pragma once
+
+class QString;
+class QStringView;
+namespace QCoro {
+    template <typename T> class AsyncGenerator;
+    template <typename T> class Task;
+}
+
+
+namespace gpt4all::ui {
+
+
+class GenerationParams;
+class ModelDescription;
+
+struct ChatResponseMetadata {
+    int nPromptTokens;
+    int nResponseTokens;
+};
+
+// TODO: implement two of these; one based on Ollama (TBD) and the other based on OpenAI (chatapi.h)
+class ChatLLMInstance {
+public:
+    virtual ~ChatLLMInstance() = 0;
+
+    virtual auto description() const -> const ModelDescription * = 0;
+    virtual auto preload() -> QCoro::Task<void> = 0;
+    virtual auto generate(QStringView prompt, const GenerationParams &params, /*out*/ ChatResponseMetadata &metadata)
+        -> QCoro::AsyncGenerator<QString> = 0;
+};
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_description.cpp b/gpt4all-chat/src/llmodel_description.cpp
new file mode 100644
index 00000000..2dd5804f
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_description.cpp
@@ -0,0 +1,22 @@
+#include "llmodel_description.h"
+
+#include "llmodel_chat.h"
+#include "llmodel_provider.h"
+
+
+namespace gpt4all::ui {
+
+
+auto ModelDescription::newInstance(QNetworkAccessManager *nam) const -> std::unique_ptr<ChatLLMInstance>
+{ return std::unique_ptr<ChatLLMInstance>(newInstanceImpl(nam)); }
+
+bool operator==(const ModelDescription &a, const ModelDescription &b)
+{
+    if (typeid(a) != typeid(b))
+        return false;
+
+    return *a.provider() == *b.provider() && a.key() == b.key();
+}
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_description.h b/gpt4all-chat/src/llmodel_description.h
new file mode 100644
index 00000000..ae9b7c14
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_description.h
@@ -0,0 +1,40 @@
+#pragma once
+
+#include <QObject>
+#include <QVariant>
+
+#include <memory>
+
+class QNetworkAccessManager;
+
+
+namespace gpt4all::ui {
+
+
+class ChatLLMInstance;
+class ModelProvider;
+
+// TODO: implement shared_from_this guidance for restricted construction
+class ModelDescription : public std::enable_shared_from_this<ModelDescription> {
+    Q_GADGET
+    Q_PROPERTY(const ModelProvider *provider READ provider CONSTANT)
+    Q_PROPERTY(QVariant             key      READ key      CONSTANT)
+
+public:
+    virtual ~ModelDescription() noexcept = 0;
+
+    // getters
+    [[nodiscard]] virtual auto     provider() const -> const ModelProvider * = 0;
+    [[nodiscard]] virtual QVariant key     () const = 0;
+
+    /// create an instance to chat with
+    [[nodiscard]] auto newInstance(QNetworkAccessManager *nam) const -> std::unique_ptr<ChatLLMInstance>;
+
+    friend bool operator==(const ModelDescription &a, const ModelDescription &b);
+
+protected:
+    [[nodiscard]] virtual auto newInstanceImpl(QNetworkAccessManager *nam) const -> ChatLLMInstance * = 0;
+};
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_ollama.cpp b/gpt4all-chat/src/llmodel_ollama.cpp
new file mode 100644
index 00000000..4a38d5a8
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_ollama.cpp
@@ -0,0 +1,82 @@
+#include "llmodel_ollama.h"
+
+#include <QCoro/QCoroAsyncGenerator>
+#include <QCoro/QCoroTask>
+
+using namespace Qt::Literals::StringLiterals;
+
+
+namespace gpt4all::ui {
+
+
+void OllamaGenerationParams::parseInner(QMap<GenerationParam, QVariant> &values)
+{
+    tryParseValue(values, GenerationParam::NPredict, &OllamaGenerationParams::n_predict);
+}
+
+auto OllamaGenerationParams::toMap() const -> QMap<QLatin1StringView, QVariant>
+{
+    return {
+        { "n_predict"_L1, n_predict },
+    };
+}
+
+auto OllamaProvider::supportedGenerationParams() const -> QSet<GenerationParam>
+{
+    using enum GenerationParam;
+    return { NPredict };
+}
+
+auto OllamaProvider::makeGenerationParams(const QMap<GenerationParam, QVariant> &values) const
+    -> OllamaGenerationParams *
+{ return new OllamaGenerationParams(values); }
+
+/// load
+OllamaProviderCustom::OllamaProviderCustom(std::shared_ptr<ProviderStore> store, QUuid id)
+    : ModelProvider(std::move(id))
+    , ModelProviderCustom(std::move(store))
+{ load(); }
+
+/// create
+OllamaProviderCustom::OllamaProviderCustom(std::shared_ptr<ProviderStore> store, QString name, QUrl baseUrl)
+    : ModelProvider(std::move(name), std::move(baseUrl))
+    , ModelProviderCustom(std::move(store))
+{
+    auto data = m_store->create(m_name, m_baseUrl);
+    if (!data)
+        data.error().raise();
+    m_id = (*data)->id;
+}
+
+OllamaModelDescription::OllamaModelDescription(std::shared_ptr<const OllamaProvider> provider, QByteArray modelHash)
+    : m_provider(std::move(provider))
+    , m_modelHash(std::move(modelHash))
+    {}
+
+auto OllamaModelDescription::newInstance(QNetworkAccessManager *nam) const -> std::unique_ptr<OllamaChatModel>
+{ return std::unique_ptr<OllamaChatModel>(&dynamic_cast<OllamaChatModel &>(*newInstanceImpl(nam))); }
+
+auto OllamaModelDescription::newInstanceImpl(QNetworkAccessManager *nam) const -> ChatLLMInstance *
+{ return new OllamaChatModel({ shared_from_this(), this }, nam); }
+
+OllamaChatModel::OllamaChatModel(std::shared_ptr<const OllamaModelDescription> description, QNetworkAccessManager *nam)
+    : m_description(std::move(description))
+    , m_nam(nam)
+    {}
+
+auto OllamaChatModel::preload() -> QCoro::Task<>
+{
+    // TODO: implement
+    co_return;
+}
+
+auto OllamaChatModel::generate(QStringView prompt, const GenerationParams &params,
+                               /*out*/ ChatResponseMetadata &metadata)
+    -> QCoro::AsyncGenerator<QString>
+{
+    // TODO: implement
+    co_yield QStringLiteral("(TODO: response from ollama)");
+}
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_ollama.h b/gpt4all-chat/src/llmodel_ollama.h
new file mode 100644
index 00000000..8a95565b
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_ollama.h
@@ -0,0 +1,122 @@
+#pragma once
+
+#include "llmodel_chat.h"
+#include "llmodel_description.h"
+#include "llmodel_provider.h"
+
+#include <QByteArray>
+#include <QLatin1StringView> // IWYU pragma: keep
+#include <QObject>
+#include <QString>
+#include <QUrl>
+#include <QVariant>
+#include <QtTypes> // IWYU pragma: keep
+
+class QNetworkAccessManager;
+template <typename Key, typename T> class QMap;
+template <typename T> class QSet;
+
+
+namespace gpt4all::ui {
+
+
+class OllamaChatModel;
+
+struct OllamaGenerationParamsData {
+    uint n_predict;
+    // TODO(jared): include ollama-specific generation params
+};
+
+class OllamaGenerationParams : public GenerationParams, public OllamaGenerationParamsData {
+public:
+    explicit OllamaGenerationParams(QMap<GenerationParam, QVariant> values) { parse(std::move(values)); }
+    auto toMap() const -> QMap<QLatin1StringView, QVariant> override;
+    bool isNoop() const override { return !n_predict; }
+
+protected:
+    void parseInner(QMap<GenerationParam, QVariant> &values) override;
+};
+
+class OllamaProvider : public QObject, public virtual ModelProvider {
+    Q_OBJECT
+
+public:
+    ~OllamaProvider() noexcept override = 0;
+
+          QObject *asQObject()       override { return this; }
+    const QObject *asQObject() const override { return this; }
+
+    auto supportedGenerationParams() const -> QSet<GenerationParam> override;
+    auto makeGenerationParams(const QMap<GenerationParam, QVariant> &values) const -> OllamaGenerationParams * override;
+};
+
+class OllamaProviderBuiltin : public ModelProviderBuiltin, public OllamaProvider {
+    Q_GADGET
+
+public:
+    /// Create a new built-in Ollama provider (transient).
+    explicit OllamaProviderBuiltin(QUuid id, QString name, QUrl baseUrl)
+        : ModelProvider(std::move(id), std::move(name), std::move(baseUrl)) {}
+};
+
+class OllamaProviderCustom final : public OllamaProvider, public ModelProviderCustom {
+    Q_OBJECT
+
+public:
+    /// Load an existing OllamaProvider from disk.
+    explicit OllamaProviderCustom(std::shared_ptr<ProviderStore> store, QUuid id);
+
+    /// Create a new OllamaProvider on disk.
+    explicit OllamaProviderCustom(std::shared_ptr<ProviderStore> store, QString name, QUrl baseUrl);
+
+Q_SIGNALS:
+    void nameChanged   (const QString &value);
+    void baseUrlChanged(const QUrl    &value);
+
+protected:
+    auto asData() -> ModelProviderData override
+    { return { m_id, ProviderType::ollama, m_name, m_baseUrl, {} }; }
+};
+
+class OllamaModelDescription : public ModelDescription {
+    Q_GADGET
+    Q_PROPERTY(QByteArray modelHash READ modelHash CONSTANT)
+
+public:
+    explicit OllamaModelDescription(std::shared_ptr<const OllamaProvider> provider, QByteArray modelHash);
+
+    // getters
+    [[nodiscard]] auto              provider () const -> const OllamaProvider * override { return m_provider.get(); }
+    [[nodiscard]] QVariant          key      () const                           override { return m_modelHash;      }
+    [[nodiscard]] const QByteArray &modelHash() const                                    { return m_modelHash;      }
+
+    [[nodiscard]] auto newInstance(QNetworkAccessManager *nam) const -> std::unique_ptr<OllamaChatModel>;
+
+protected:
+    [[nodiscard]] auto newInstanceImpl(QNetworkAccessManager *nam) const -> ChatLLMInstance * override;
+
+private:
+    std::shared_ptr<const OllamaProvider> m_provider;
+    QByteArray                            m_modelHash;
+};
+
+class OllamaChatModel : public ChatLLMInstance {
+public:
+    explicit OllamaChatModel(std::shared_ptr<const OllamaModelDescription> description, QNetworkAccessManager *nam);
+
+    auto description() const -> const OllamaModelDescription * override
+    { return m_description.get(); }
+
+    auto preload() -> QCoro::Task<void> override;
+
+    auto generate(QStringView prompt, const GenerationParams &params, /*out*/ ChatResponseMetadata &metadata)
+        -> QCoro::AsyncGenerator<QString> override;
+
+private:
+    std::shared_ptr<const OllamaModelDescription> m_description;
+    // TODO: implement generate using Ollama backend
+    QNetworkAccessManager                         *m_nam;
+};
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel/openai.cpp b/gpt4all-chat/src/llmodel_openai.cpp
similarity index 59%
rename from gpt4all-chat/src/llmodel/openai.cpp
rename to gpt4all-chat/src/llmodel_openai.cpp
index 32a0b58f..6be8f9bf 100644
--- a/gpt4all-chat/src/llmodel/openai.cpp
+++ b/gpt4all-chat/src/llmodel_openai.cpp
@@ -1,4 +1,4 @@
-#include "openai.h"
+#include "llmodel_openai.h"
 
 #include "mysettings.h"
 #include "utils.h"
@@ -6,29 +6,34 @@
 #include <QCoro/QCoroAsyncGenerator> // IWYU pragma: keep
 #include <QCoro/QCoroNetworkReply> // IWYU pragma: keep
 #include <fmt/format.h>
-#include <gpt4all-backend/formatters.h>
-#include <gpt4all-backend/generation-params.h>
+#include <gpt4all-backend/formatters.h> // IWYU pragma: keep
 #include <gpt4all-backend/rest.h>
 
+#include <QAnyStringView>
 #include <QByteArray>
 #include <QJsonArray>
 #include <QJsonDocument>
 #include <QJsonObject>
 #include <QJsonValue>
-#include <QLatin1String>
-#include <QNetworkAccessManager>
+#include <QList>
+#include <QMap>
+#include <QMetaEnum>
+#include <QNetworkReply>
 #include <QNetworkRequest>
 #include <QRestAccessManager>
 #include <QRestReply>
+#include <QSet>
 #include <QStringView>
-#include <QUrl>
 #include <QUtf8StringView> // IWYU pragma: keep
 #include <QVariant>
 #include <QXmlStreamReader>
-#include <Qt>
+#include <QtAssert>
 
+#include <coroutine>
 #include <expected>
+#include <memory>
 #include <optional>
+#include <stdexcept>
 #include <utility>
 
 using namespace Qt::Literals::StringLiterals;
@@ -63,24 +68,72 @@ static auto processRespLine(const QByteArray &line) -> std::optional<QString>
 namespace gpt4all::ui {
 
 
-void OpenaiModelDescription::setDisplayName(QString value)
+void OpenaiGenerationParams::parseInner(QMap<GenerationParam, QVariant> &values)
 {
-    if (m_displayName != value) {
-        m_displayName = std::move(value);
-        emit displayNameChanged(m_displayName);
-    }
+    tryParseValue(values, GenerationParam::NPredict,    &OpenaiGenerationParams::n_predict  );
+    tryParseValue(values, GenerationParam::Temperature, &OpenaiGenerationParams::temperature);
+    tryParseValue(values, GenerationParam::TopP,        &OpenaiGenerationParams::top_p      );
 }
 
-void OpenaiModelDescription::setModelName(QString value)
+auto OpenaiGenerationParams::toMap() const -> QMap<QLatin1StringView, QVariant>
 {
-    if (m_modelName != value) {
-        m_modelName = std::move(value);
-        emit modelNameChanged(m_modelName);
-    }
+    return {
+        {  "max_completion_tokens"_L1,  n_predict   },
+        {  "temperature"_L1,            temperature },
+        {  "top_p"_L1,                  top_p       },
+    };
 }
 
-OpenaiLLModel::OpenaiLLModel(OpenaiConnectionDetails connDetails, QNetworkAccessManager *nam)
-    : m_connDetails(std::move(connDetails))
+auto OpenaiProvider::supportedGenerationParams() const -> QSet<GenerationParam>
+{
+    using enum GenerationParam;
+    return { NPredict, Temperature, TopP };
+}
+
+auto OpenaiProvider::makeGenerationParams(const QMap<GenerationParam, QVariant> &values) const
+    -> OpenaiGenerationParams *
+{ return new OpenaiGenerationParams(values); }
+
+OpenaiProviderBuiltin::OpenaiProviderBuiltin(QUuid id, QString name, QUrl baseUrl, QString apiKey)
+    : ModelProvider(std::move(id), std::move(name), std::move(baseUrl))
+    , OpenaiProvider(std::move(apiKey))
+    {}
+
+/// load
+OpenaiProviderCustom::OpenaiProviderCustom(std::shared_ptr<ProviderStore> store, QUuid id)
+    : ModelProvider(std::move(id))
+    , ModelProviderCustom(std::move(store))
+{
+    auto &details = load();
+    m_apiKey = std::get<OpenaiProviderDetails>(details).api_key;
+}
+
+/// create
+OpenaiProviderCustom::OpenaiProviderCustom(std::shared_ptr<ProviderStore> store, QString name, QUrl baseUrl,
+                                           QString apiKey)
+    : ModelProvider(std::move(name), std::move(baseUrl))
+    , ModelProviderCustom(std::move(store))
+    , OpenaiProvider(std::move(apiKey))
+{
+    auto data = m_store->create(m_name, m_baseUrl, m_apiKey);
+    if (!data)
+        data.error().raise();
+    m_id = (*data)->id;
+}
+
+OpenaiModelDescription::OpenaiModelDescription(std::shared_ptr<const OpenaiProvider> provider, QString modelName)
+    : m_provider(std::move(provider))
+    , m_modelName(std::move(modelName))
+    {}
+
+auto OpenaiModelDescription::newInstance(QNetworkAccessManager *nam) const -> std::unique_ptr<OpenaiChatModel>
+{ return std::unique_ptr<OpenaiChatModel>(&dynamic_cast<OpenaiChatModel &>(*newInstanceImpl(nam))); }
+
+auto OpenaiModelDescription::newInstanceImpl(QNetworkAccessManager *nam) const -> ChatLLMInstance *
+{ return new OpenaiChatModel({ shared_from_this(), this }, nam); }
+
+OpenaiChatModel::OpenaiChatModel(std::shared_ptr<const OpenaiModelDescription> description, QNetworkAccessManager *nam)
+    : m_description(std::move(description))
     , m_nam(nam)
     {}
 
@@ -159,21 +212,22 @@ static auto parsePrompt(QXmlStreamReader &xml) -> std::expected<QJsonArray, QStr
     }
 }
 
-auto OpenaiLLModel::chat(QStringView prompt, const backend::GenerationParams &params,
-                         /*out*/ ChatResponseMetadata &metadata) -> QCoro::AsyncGenerator<QString>
+auto preload() -> QCoro::Task<>
+{ co_return; /* not supported -> no-op */ }
+
+auto OpenaiChatModel::generate(QStringView prompt, const GenerationParams &params,
+                               /*out*/ ChatResponseMetadata &metadata) -> QCoro::AsyncGenerator<QString>
 {
     auto *mySettings = MySettings::globalInstance();
 
-    if (!params.n_predict)
+    if (params.isNoop())
         co_return; // nothing requested
 
     auto reqBody = makeJsonObject({
-        { "model"_L1,                 m_connDetails.modelName  },
-        { "max_completion_tokens"_L1, qint64(params.n_predict) },
-        { "stream"_L1,                true                     },
-        { "temperature"_L1,           params.temperature       },
-        { "top_p"_L1,                 params.top_p             },
+        { "model"_L1,  m_description->modelName() },
+        { "stream"_L1, true                       },
     });
+    extend(reqBody, params.toMap());
 
     // conversation history
     {
@@ -184,9 +238,10 @@ auto OpenaiLLModel::chat(QStringView prompt, const backend::GenerationParams &pa
         reqBody.insert("messages"_L1, *messages);
     }
 
-    QNetworkRequest request(m_connDetails.baseUrl.resolved(QUrl("/v1/chat/completions")));
+    auto &provider = *m_description->provider();
+    QNetworkRequest request(provider.baseUrl().resolved(QUrl("/v1/chat/completions")));
     request.setHeader(QNetworkRequest::UserAgentHeader, mySettings->userAgent());
-    request.setRawHeader("authorization", u"Bearer %1"_s.arg(m_connDetails.apiKey).toUtf8());
+    request.setRawHeader("authorization", u"Bearer %1"_s.arg(provider.apiKey()).toUtf8());
 
     QRestAccessManager restNam(m_nam);
     std::unique_ptr<QNetworkReply> reply(restNam.post(request, QJsonDocument(reqBody)));
diff --git a/gpt4all-chat/src/llmodel_openai.h b/gpt4all-chat/src/llmodel_openai.h
new file mode 100644
index 00000000..50dba4d7
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_openai.h
@@ -0,0 +1,139 @@
+#pragma once
+
+#include "llmodel_chat.h"
+#include "llmodel_description.h"
+#include "llmodel_provider.h"
+
+#include <QLatin1StringView> // IWYU pragma: keep
+#include <QObject> // IWYU pragma: keep
+#include <QString>
+#include <QUrl>
+#include <QVariant>
+#include <QtTypes> // IWYU pragma: keep
+
+#include <memory>
+#include <utility>
+
+class QNetworkAccessManager;
+template <typename Key, typename T> class QMap;
+template <typename T> class QSet;
+
+
+namespace gpt4all::ui {
+
+
+class OpenaiChatModel;
+
+struct OpenaiGenerationParamsData {
+    uint  n_predict;
+    float temperature;
+    float top_p;
+};
+
+class OpenaiGenerationParams : public GenerationParams, public OpenaiGenerationParamsData {
+public:
+    explicit OpenaiGenerationParams(QMap<GenerationParam, QVariant> values) { parse(std::move(values)); }
+    auto toMap() const -> QMap<QLatin1StringView, QVariant> override;
+    bool isNoop() const override { return !n_predict; }
+
+protected:
+    void parseInner(QMap<GenerationParam, QVariant> &values) override;
+};
+
+class OpenaiProvider : public QObject, public virtual ModelProvider {
+    Q_OBJECT
+
+protected:
+    explicit OpenaiProvider() = default; // custom
+    explicit OpenaiProvider(QString apiKey) // built-in
+        : m_apiKey(std::move(apiKey))
+        {}
+public:
+    ~OpenaiProvider() noexcept override = 0;
+
+          QObject *asQObject()       override { return this; }
+    const QObject *asQObject() const override { return this; }
+
+    [[nodiscard]] const QString &apiKey() const { return m_apiKey; }
+
+    auto supportedGenerationParams() const -> QSet<GenerationParam> override;
+    auto makeGenerationParams(const QMap<GenerationParam, QVariant> &values) const -> OpenaiGenerationParams * override;
+
+protected:
+    QString m_apiKey;
+};
+
+class OpenaiProviderBuiltin : public ModelProviderBuiltin, public OpenaiProvider {
+    Q_GADGET
+    Q_PROPERTY(QString apiKey READ apiKey CONSTANT)
+
+public:
+    /// Create a new built-in OpenAI provider (transient).
+    explicit OpenaiProviderBuiltin(QUuid id, QString name, QUrl baseUrl, QString apiKey);
+};
+
+class OpenaiProviderCustom final : public OpenaiProvider, public ModelProviderCustom {
+    Q_OBJECT
+
+    Q_PROPERTY(QString apiKey READ apiKey WRITE setApiKey NOTIFY apiKeyChanged)
+
+public:
+    /// Load an existing OpenaiProvider from disk.
+    explicit OpenaiProviderCustom(std::shared_ptr<ProviderStore> store, QUuid id);
+
+    /// Create a new OpenaiProvider on disk.
+    explicit OpenaiProviderCustom(std::shared_ptr<ProviderStore> store, QString name, QUrl baseUrl, QString apiKey);
+
+    void setApiKey(QString value) { setMemberProp<QString>(&OpenaiProviderCustom::m_apiKey, "apiKey", std::move(value)); }
+
+Q_SIGNALS:
+    void nameChanged   (const QString &value);
+    void baseUrlChanged(const QUrl    &value);
+    void apiKeyChanged (const QString &value);
+
+protected:
+    auto asData() -> ModelProviderData override
+    { return { m_id, ProviderType::openai, m_name, m_baseUrl, OpenaiProviderDetails { m_apiKey } }; }
+};
+
+class OpenaiModelDescription : public ModelDescription {
+    Q_GADGET
+    Q_PROPERTY(QString modelName READ modelName CONSTANT)
+
+public:
+    explicit OpenaiModelDescription(std::shared_ptr<const OpenaiProvider> provider, QString modelName);
+
+    // getters
+    [[nodiscard]] auto           provider () const -> const OpenaiProvider * override { return m_provider.get(); }
+    [[nodiscard]] QVariant       key      () const                           override { return m_modelName;      }
+    [[nodiscard]] const QString &modelName() const                                    { return m_modelName;      }
+
+    [[nodiscard]] auto newInstance(QNetworkAccessManager *nam) const -> std::unique_ptr<OpenaiChatModel>;
+
+protected:
+    [[nodiscard]] auto newInstanceImpl(QNetworkAccessManager *nam) const -> ChatLLMInstance * override;
+
+private:
+    std::shared_ptr<const OpenaiProvider> m_provider;
+    QString                               m_modelName;
+};
+
+class OpenaiChatModel : public ChatLLMInstance {
+public:
+    explicit OpenaiChatModel(std::shared_ptr<const OpenaiModelDescription> description, QNetworkAccessManager *nam);
+
+    auto description() const -> const OpenaiModelDescription * override
+    { return m_description.get(); }
+
+    auto preload() -> QCoro::Task<void> override;
+
+    auto generate(QStringView prompt, const GenerationParams &params, /*out*/ ChatResponseMetadata &metadata)
+        -> QCoro::AsyncGenerator<QString> override;
+
+private:
+    std::shared_ptr<const OpenaiModelDescription>  m_description;
+    QNetworkAccessManager                         *m_nam;
+};
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_provider.cpp b/gpt4all-chat/src/llmodel_provider.cpp
new file mode 100644
index 00000000..9ae83410
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_provider.cpp
@@ -0,0 +1,139 @@
+#include "llmodel_provider.h"
+
+#include "mysettings.h"
+
+#include <fmt/format.h>
+#include <gpt4all-backend/formatters.h> // IWYU pragma: keep
+
+#include <QModelIndex> // IWYU pragma: keep
+#include <QVariant>
+
+namespace fs = std::filesystem;
+
+
+namespace gpt4all::ui {
+
+
+void GenerationParams::parse(QMap<GenerationParam, QVariant> values)
+{
+    parseInner(values);
+    if (!values.isEmpty()) {
+        auto gparamsMeta = QMetaEnum::fromType<GenerationParam>();
+        throw std::invalid_argument(fmt::format(
+            " unsupported param: {}", gparamsMeta.valueToKey(int(values.keys().constFirst()))
+        ));
+    }
+}
+
+QVariant GenerationParams::tryParseValue(QMap<GenerationParam, QVariant> &values, GenerationParam key,
+                                         const QMetaType &type)
+{
+    auto value = values.take(key);
+    if (value.isValid() && !value.canConvert(type)) {
+        auto gparamsMeta = QMetaEnum::fromType<GenerationParam>();
+        throw std::invalid_argument(fmt::format(
+            "expected {} of type {}, got {}", gparamsMeta.valueToKey(int(key)), type.name(), value.typeName()
+        ));
+    }
+    return value;
+}
+
+ModelProviderCustom::~ModelProviderCustom() noexcept
+{
+    if (auto res = m_store->release(m_id); !res)
+        res.error().raise(); // should not happen - will terminate program
+}
+
+auto ModelProviderCustom::load() -> const ModelProviderData::Details &
+{
+    auto data = m_store->acquire(m_id);
+    if (!data)
+        data.error().raise();
+    m_name    = (*data)->name;
+    m_baseUrl = (*data)->base_url;
+    return (*data)->details;
+}
+
+ProviderRegistry::ProviderRegistry(fs::path path)
+    : m_store(std::move(path))
+{
+    auto *mysettings = MySettings::globalInstance();
+    connect(mysettings, &MySettings::modelPathChanged, this, &ProviderRegistry::onModelPathChanged);
+}
+
+Q_INVOKABLE void ProviderRegistry::registerBuiltinProvider(ModelProviderBuiltin *provider)
+{
+    auto [_, unique] = m_providers.emplace(provider->id(), provider->asQObject());
+    if (!unique)
+        qWarning() << "ignoring duplicate provider:" << provider->id();
+}
+
+[[nodiscard]]
+bool ProviderRegistry::registerCustomProvider(std::unique_ptr<ModelProviderCustom> provider)
+{
+    auto [_, unique] = m_providers.emplace(provider->id(), provider->asQObject());
+    if (unique) {
+        m_customProviders.push_back(std::move(provider));
+        emit customProviderAdded(m_customProviders.size() - 1);
+    }
+    return unique;
+}
+
+fs::path ProviderRegistry::getSubdir()
+{
+    auto *mysettings = MySettings::globalInstance();
+    return toFSPath(mysettings->modelPath()) / "providers";
+}
+
+void ProviderRegistry::onModelPathChanged()
+{
+    auto path = getSubdir();
+    if (path != m_store.path()) {
+        emit aboutToBeCleared();
+        m_customProviders.clear(); // delete custom providers to release store locks
+        if (auto res = m_store.setPath(path); !res)
+            res.error().raise(); // should not happen
+    }
+}
+
+CustomProviderList::CustomProviderList(QPointer<ProviderRegistry> registry)
+    : m_registry(std::move(registry))
+    , m_size(m_registry->customProviderCount())
+{
+    connect(m_registry, &ProviderRegistry::customProviderAdded, this, &CustomProviderList::onCustomProviderAdded);
+    connect(m_registry, &ProviderRegistry::aboutToBeCleared, this, &CustomProviderList::onAboutToBeCleared,
+            Qt::DirectConnection);
+}
+
+QVariant CustomProviderList::data(const QModelIndex &index, int role) const
+{
+    if (index.isValid() && index.row() < rowCount() && role == Qt::DisplayRole)
+        return QVariant::fromValue(m_registry->customProviderAt(index.row()));
+    return {};
+}
+
+void CustomProviderList::onCustomProviderAdded(size_t index)
+{
+    beginInsertRows({}, m_size, m_size);
+    m_size++;
+    endInsertRows();
+}
+
+void CustomProviderList::onAboutToBeCleared()
+{
+    beginResetModel();
+    m_size = 0;
+    endResetModel();
+}
+
+bool CustomProviderListSort::lessThan(const QModelIndex &left, const QModelIndex &right) const
+{
+    auto *leftData  = sourceModel()->data(left ).value<ModelProviderCustom *>();
+    auto *rightData = sourceModel()->data(right).value<ModelProviderCustom *>();
+    if (leftData && rightData)
+        return QString::localeAwareCompare(leftData->name(), rightData->name()) < 0;
+    return true;
+}
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/llmodel_provider.h b/gpt4all-chat/src/llmodel_provider.h
new file mode 100644
index 00000000..5b8c3595
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_provider.h
@@ -0,0 +1,199 @@
+#pragma once
+
+#include "store_provider.h"
+
+#include "utils.h" // IWYU pragma: keep
+
+#include <QAbstractListModel>
+#include <QObject>
+#include <QPointer>
+#include <QQmlEngine> // IWYU pragma: keep
+#include <QSortFilterProxyModel>
+#include <QString>
+#include <QUrl>
+#include <QUuid>
+#include <QtPreprocessorSupport>
+
+#include <cstddef>
+#include <filesystem>
+#include <memory>
+#include <string_view>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+class QJSEngine;
+
+
+namespace gpt4all::ui {
+
+
+Q_NAMESPACE
+
+enum class GenerationParam  {
+    NPredict,
+    Temperature,
+    TopP,
+    TopK,
+    MinP,
+    RepeatPenalty,
+    RepeatLastN,
+};
+Q_ENUM_NS(GenerationParam)
+
+class GenerationParams {
+public:
+    virtual ~GenerationParams() noexcept = 0;
+
+    virtual QMap<QLatin1StringView, QVariant> toMap() const = 0;
+    virtual bool isNoop() const = 0;
+
+protected:
+    void parse(QMap<GenerationParam, QVariant> values);
+    virtual void parseInner(QMap<GenerationParam, QVariant> &values) = 0;
+
+    static QVariant tryParseValue(QMap<GenerationParam, QVariant> &values, GenerationParam key, const QMetaType &type);
+
+    template <typename T, typename S, typename C>
+    void tryParseValue(this S &self, QMap<GenerationParam, QVariant> &values, GenerationParam key, T C::* dest);
+};
+
+class ModelProvider {
+    Q_GADGET
+    Q_PROPERTY(QUuid id READ id CONSTANT)
+
+protected:
+    explicit ModelProvider(QUuid id) // load
+        : m_id(std::move(id)) {}
+    explicit ModelProvider(QUuid id, QString name, QUrl baseUrl) // create built-in
+        : m_id(std::move(id)), m_name(std::move(name)), m_baseUrl(std::move(baseUrl)) {}
+    explicit ModelProvider(QString name, QUrl baseUrl) // create custom
+        : m_name(std::move(name)), m_baseUrl(std::move(baseUrl)) {}
+
+public:
+    virtual ~ModelProvider() noexcept = 0;
+
+    virtual       QObject *asQObject() = 0;
+    virtual const QObject *asQObject() const = 0;
+
+    // getters
+    [[nodiscard]] const QUuid   &id     () const { return m_id;      }
+    [[nodiscard]] const QString &name   () const { return m_name;    }
+    [[nodiscard]] const QUrl    &baseUrl() const { return m_baseUrl; }
+
+    virtual auto supportedGenerationParams() const -> QSet<GenerationParam> = 0;
+    virtual auto makeGenerationParams(const QMap<GenerationParam, QVariant> &values) const -> GenerationParams * = 0;
+
+    friend bool operator==(const ModelProvider &a, const ModelProvider &b)
+    { return a.m_id == b.m_id; }
+
+protected:
+    QUuid   m_id;
+    QString m_name;
+    QUrl    m_baseUrl;
+};
+
+class ModelProviderBuiltin : public virtual ModelProvider {
+    Q_GADGET
+    Q_PROPERTY(QString name    READ name    CONSTANT)
+    Q_PROPERTY(QUrl    baseUrl READ baseUrl CONSTANT)
+
+public:
+    ~ModelProviderBuiltin() noexcept override = 0;
+};
+
+class ModelProviderCustom : public virtual ModelProvider {
+    Q_GADGET
+    Q_PROPERTY(QString name    READ name    WRITE setName    NOTIFY nameChanged   )
+    Q_PROPERTY(QUrl    baseUrl READ baseUrl WRITE setBaseUrl NOTIFY baseUrlChanged)
+
+protected:
+    explicit ModelProviderCustom(std::shared_ptr<ProviderStore> store)
+        : m_store(std::move(store)) {}
+
+public:
+    ~ModelProviderCustom() noexcept override;
+
+    // setters
+    void setName   (QString value) { setMemberProp<QString>(&ModelProviderCustom::m_name,    "name",    std::move(value)); }
+    void setBaseUrl(QUrl    value) { setMemberProp<QUrl   >(&ModelProviderCustom::m_baseUrl, "baseUrl", std::move(value)); }
+
+protected:
+    virtual auto load() -> const ModelProviderData::Details &;
+    virtual auto asData() -> ModelProviderData = 0;
+
+    template <typename T, typename S, typename C>
+    void setMemberProp(this S &self, T C::* member, std::string_view name, T value);
+
+    std::shared_ptr<ProviderStore> m_store;
+};
+
+class ProviderRegistry : public QObject {
+    Q_OBJECT
+    QML_ELEMENT
+    QML_SINGLETON
+
+protected:
+    explicit ProviderRegistry(std::filesystem::path path);
+
+public:
+    static ProviderRegistry *create(QQmlEngine *, QJSEngine *) { return new ProviderRegistry(getSubdir()); }
+    Q_INVOKABLE   void registerBuiltinProvider(ModelProviderBuiltin *provider);
+    [[nodiscard]] bool registerCustomProvider (std::unique_ptr<ModelProviderCustom> provider);
+
+    size_t customProviderCount() const
+    { return m_customProviders.size(); }
+    auto customProviderAt(size_t i) const -> const ModelProviderCustom *
+    { return m_customProviders.at(i).get(); }
+    auto operator[](const QUuid &id) -> ModelProviderCustom *
+    { return &dynamic_cast<ModelProviderCustom &>(*m_providers.at(id)); }
+
+Q_SIGNALS:
+    void customProviderAdded(size_t index);
+    void aboutToBeCleared();
+
+private:
+    static auto getSubdir() -> std::filesystem::path;
+
+private Q_SLOTS:
+    void onModelPathChanged();
+
+private:
+    ProviderStore                                     m_store;
+    std::unordered_map<QUuid, QPointer<QObject>>      m_providers;
+    std::vector<std::unique_ptr<ModelProviderCustom>> m_customProviders;
+};
+
+class CustomProviderList : public QAbstractListModel {
+    Q_OBJECT
+    QML_ELEMENT
+
+protected:
+    explicit CustomProviderList(QPointer<ProviderRegistry> registry);
+
+public:
+    int rowCount(const QModelIndex &parent = {}) const override
+    { Q_UNUSED(parent) return int(m_size); }
+    QVariant data(const QModelIndex &index, int role) const override;
+
+private Q_SLOTS:
+    void onCustomProviderAdded(size_t index);
+    void onAboutToBeCleared();
+
+private:
+    QPointer<ProviderRegistry> m_registry;
+    size_t                     m_size;
+};
+
+class CustomProviderListSort : public QSortFilterProxyModel {
+    Q_OBJECT
+
+protected:
+    bool lessThan(const QModelIndex &left, const QModelIndex &right) const override;
+};
+
+
+} // namespace gpt4all::ui
+
+
+#include "llmodel_provider.inl" // IWYU pragma: export
diff --git a/gpt4all-chat/src/llmodel_provider.inl b/gpt4all-chat/src/llmodel_provider.inl
new file mode 100644
index 00000000..a707f6b2
--- /dev/null
+++ b/gpt4all-chat/src/llmodel_provider.inl
@@ -0,0 +1,30 @@
+#include <fmt/format.h>
+
+
+namespace gpt4all::ui {
+
+
+template <typename T, typename S, typename C>
+void GenerationParams::tryParseValue(this S &self, QMap<GenerationParam, QVariant> &values, GenerationParam key,
+                                     T C::* dest)
+{
+    if (auto value = tryParseValue(values, key, QMetaType::fromType<T>()); value.isValid())
+        self.*dest = value.template value<T>();
+}
+
+template <typename T, typename S, typename C>
+void ModelProviderCustom::setMemberProp(this S &self, T C::* member, std::string_view name, T value)
+{
+    auto &mpc = static_cast<ModelProviderCustom &>(self);
+    auto &cur = self.*member;
+    if (cur != value) {
+        cur = std::move(value);
+        auto data = mpc.asData();
+        if (auto res = mpc.m_store->setData(std::move(data)); !res)
+            res.error().raise();
+        QMetaObject::invokeMethod(self.asQObject(), fmt::format("{}Changed", name).c_str(), cur);
+    }
+}
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/modellist.h b/gpt4all-chat/src/modellist.h
index 756c6ffe..5520b288 100644
--- a/gpt4all-chat/src/modellist.h
+++ b/gpt4all-chat/src/modellist.h
@@ -16,6 +16,7 @@
 #include <QSortFilterProxyModel>
 #include <QSslError>
 #include <QString>
+#include <QUuid>
 #include <QVariant>
 #include <QVector> // IWYU pragma: keep
 #include <Qt>
@@ -27,6 +28,7 @@
 // IWYU pragma: no_forward_declare QObject
 // IWYU pragma: no_forward_declare QSslError
 class QUrl;
+namespace gpt4all::ui { class ModelDescription; }
 
 using namespace Qt::Literals::StringLiterals;
 
@@ -75,6 +77,7 @@ private:
 struct ModelInfo {
     Q_GADGET
     Q_PROPERTY(QString id READ id WRITE setId)
+    Q_PROPERTY(const ModelDescription *modelDesc READ modelDescQt WRITE setModelDescQt)
     Q_PROPERTY(QString name READ name WRITE setName)
     Q_PROPERTY(QString filename READ filename WRITE setFilename)
     Q_PROPERTY(QString dirpath MEMBER dirpath)
@@ -137,6 +140,13 @@ public:
     QString id() const;
     void setId(const QString &id);
 
+    auto modelDesc() const -> const std::shared_ptr<const gpt4all::ui::ModelDescription> &;
+    auto modelDescQt() const -> const gpt4all::ui::ModelDescription *
+    { return modelDesc().get(); }
+
+    void setModelDesc(std::shared_ptr<const gpt4all::ui::ModelDescription> value);
+    void setModelDescQt(const gpt4all::ui::ModelDescription *); // TODO: implement
+
     QString name() const;
     void setName(const QString &name);
 
@@ -247,6 +257,7 @@ private:
     QVariant getField(QLatin1StringView name) const;
 
     QString m_id;
+    std::shared_ptr<const gpt4all::ui::ModelDescription> m_modelDesc;
     QString m_name;
     QString m_filename;
     QString m_description;
diff --git a/gpt4all-chat/src/mysettings.cpp b/gpt4all-chat/src/mysettings.cpp
index 06a77f26..dc538d90 100644
--- a/gpt4all-chat/src/mysettings.cpp
+++ b/gpt4all-chat/src/mysettings.cpp
@@ -48,7 +48,6 @@ namespace ModelSettingsKey { namespace {
 
 namespace defaults {
 
-static const int     threadCount             = std::min(4, (int32_t) std::thread::hardware_concurrency());
 static const bool    networkIsActive         = false;
 static const bool    networkUsageStatsActive = false;
 static const QString device                  = "Auto";
@@ -254,7 +253,6 @@ void MySettings::restoreApplicationDefaults()
     setChatTheme(basicDefaults.value("chatTheme").value<ChatTheme>());
     setFontSize(basicDefaults.value("fontSize").value<FontSize>());
     setDevice(defaults::device);
-    setThreadCount(defaults::threadCount);
     setSystemTray(basicDefaults.value("systemTray").toBool());
     setServerChat(basicDefaults.value("serverChat").toBool());
     setNetworkPort(basicDefaults.value("networkPort").toInt());
@@ -596,29 +594,6 @@ void MySettings::setModelSuggestedFollowUpPrompt(const ModelInfo &info, const QS
     setModelSetting("suggestedFollowUpPrompt", info, value, force, true);
 }
 
-int MySettings::threadCount() const
-{
-    int c = m_settings.value("threadCount", defaults::threadCount).toInt();
-    // The old thread setting likely left many people with 0 in settings config file, which means
-    // we should reset it to the default going forward
-    if (c <= 0)
-        c = defaults::threadCount;
-    c = std::max(c, 1);
-    c = std::min(c, QThread::idealThreadCount());
-    return c;
-}
-
-void MySettings::setThreadCount(int value)
-{
-    if (threadCount() == value)
-        return;
-
-    value = std::max(value, 1);
-    value = std::min(value, QThread::idealThreadCount());
-    m_settings.setValue("threadCount", value);
-    emit threadCountChanged();
-}
-
 bool        MySettings::systemTray() const              { return getBasicSetting("systemTray"              ).toBool(); }
 bool        MySettings::serverChat() const              { return getBasicSetting("serverChat"              ).toBool(); }
 int         MySettings::networkPort() const             { return getBasicSetting("networkPort"             ).toInt(); }
diff --git a/gpt4all-chat/src/mysettings.h b/gpt4all-chat/src/mysettings.h
index ffb7113a..29fe1866 100644
--- a/gpt4all-chat/src/mysettings.h
+++ b/gpt4all-chat/src/mysettings.h
@@ -20,6 +20,7 @@
 
 // IWYU pragma: no_forward_declare QModelIndex
 class QLocale;
+namespace gpt4all::ui { class GenerationParams; }
 
 
 namespace MySettingsEnums {
@@ -54,7 +55,6 @@ using namespace MySettingsEnums;
 class MySettings : public QObject
 {
     Q_OBJECT
-    Q_PROPERTY(int threadCount READ threadCount WRITE setThreadCount NOTIFY threadCountChanged)
     Q_PROPERTY(bool systemTray READ systemTray WRITE setSystemTray NOTIFY systemTrayChanged)
     Q_PROPERTY(bool serverChat READ serverChat WRITE setServerChat NOTIFY serverChatChanged)
     Q_PROPERTY(QString modelPath READ modelPath WRITE setModelPath NOTIFY modelPathChanged)
@@ -156,9 +156,10 @@ public:
     QString modelSuggestedFollowUpPrompt(const ModelInfo &info) const;
     Q_INVOKABLE void setModelSuggestedFollowUpPrompt(const ModelInfo &info, const QString &value, bool force = false);
 
+    // TODO: implement
+    auto modelGenParams(const ModelInfo &info) -> gpt4all::ui::GenerationParams;
+
     // Application settings
-    int threadCount() const;
-    void setThreadCount(int value);
     bool systemTray() const;
     void setSystemTray(bool value);
     bool serverChat() const;
@@ -173,10 +174,6 @@ public:
     void setFontSize(FontSize value);
     QString device();
     void setDevice(const QString &value);
-    int32_t contextLength() const;
-    void setContextLength(int32_t value);
-    int32_t gpuLayers() const;
-    void setGpuLayers(int32_t value);
     SuggestionMode suggestionMode() const;
     void setSuggestionMode(SuggestionMode value);
 
@@ -231,7 +228,6 @@ Q_SIGNALS:
     void systemMessageChanged(const ModelInfo &info, bool fromInfo = false);
     void chatNamePromptChanged(const ModelInfo &info);
     void suggestedFollowUpPromptChanged(const ModelInfo &info);
-    void threadCountChanged();
     void systemTrayChanged();
     void serverChatChanged();
     void modelPathChanged();
diff --git a/gpt4all-chat/src/server.cpp b/gpt4all-chat/src/server.cpp
index 45db49f0..42ca9c80 100644
--- a/gpt4all-chat/src/server.cpp
+++ b/gpt4all-chat/src/server.cpp
@@ -2,6 +2,7 @@
 
 #include "chat.h"
 #include "chatmodel.h"
+#include "llmodel_description.h"
 #include "modellist.h"
 #include "mysettings.h"
 
@@ -50,8 +51,10 @@
 #include <variant>
 #include <vector>
 
-using namespace std::string_literals;
 using namespace Qt::Literals::StringLiterals;
+using namespace std::string_literals;
+using namespace gpt4all;
+using namespace gpt4all::ui;
 
 //#define DEBUG
 
@@ -127,11 +130,11 @@ class BaseCompletionRequest {
 public:
     QString model; // required
     // NB: some parameters are not supported yet
-    uint max_tokens = 16;
     qint64 n = 1;
-    float temperature = 1.f;
-    float top_p = 1.f;
-    float min_p = 0.f;
+    std::optional<uint> max_tokens {};
+    std::optional<float> temperature {};
+    std::optional<float> top_p {};
+    std::optional<float> min_p {};
 
     BaseCompletionRequest() = default;
     virtual ~BaseCompletionRequest() = default;
@@ -162,7 +165,7 @@ protected:
 
         value = reqValue("max_tokens", Integer, false, /*min*/ 1);
         if (!value.isNull())
-            this->max_tokens = uint(qMin(value.toInteger(), UINT32_MAX));
+            this->max_tokens = uint(qMin(value.toInteger(), qint64(UINT32_MAX)));
 
         value = reqValue("n", Integer, false, /*min*/ 1);
         if (!value.isNull())
@@ -629,8 +632,6 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
 {
     Q_ASSERT(m_chatModel);
 
-    auto *mySettings = MySettings::globalInstance();
-
     ModelInfo modelInfo = ModelList::globalInstance()->defaultModelInfo();
     const QList<ModelInfo> modelList = ModelList::globalInstance()->selectableModelList();
     for (const ModelInfo &info : modelList) {
@@ -662,22 +663,25 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
         return makeError(QHttpServerResponder::StatusCode::InternalServerError);
     }
 
+    std::unique_ptr<GenerationParams> genParams;
+    {
+        using enum GenerationParam;
+        QMap<GenerationParam, QVariant> values;
+        if (auto v = request.max_tokens ) values.insert(NPredict,    *v);
+        if (auto v = request.temperature) values.insert(Temperature, *v);
+        if (auto v = request.top_p      ) values.insert(TopP,        *v);
+        if (auto v = request.min_p      ) values.insert(MinP,        *v);
+        try {
+            genParams.reset(modelDescription()->makeGenerationParams(values));
+        } catch (const std::exception &e) {
+            throw InvalidRequestError(e.what());
+        }
+    }
+
     // add prompt/response items to GUI
     m_chatModel->appendPrompt(request.prompt);
     m_chatModel->appendResponse();
 
-    // FIXME(jared): taking parameters from the UI inhibits reproducibility of results
-    backend::GenerationParams genParams {
-        .n_predict      = request.max_tokens,
-        .top_k          = mySettings->modelTopK(modelInfo),
-        .top_p          = request.top_p,
-        .min_p          = request.min_p,
-        .temp           = request.temperature,
-        .n_batch        = mySettings->modelPromptBatchSize(modelInfo),
-        .repeat_penalty = float(mySettings->modelRepeatPenalty(modelInfo)),
-        .repeat_last_n  = mySettings->modelRepeatPenaltyTokens(modelInfo),
-    };
-
     auto promptUtf8 = request.prompt.toUtf8();
     int promptTokens = 0;
     int responseTokens = 0;
@@ -686,7 +690,7 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
         PromptResult result;
         try {
             result = promptInternal(std::string_view(promptUtf8.cbegin(), promptUtf8.cend()),
-                                    genParams,
+                                    *genParams,
                                     /*usedLocalDocs*/ false);
         } catch (const std::exception &e) {
             m_chatModel->setResponseValue(e.what());
@@ -733,8 +737,6 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
 auto Server::handleChatRequest(const ChatRequest &request)
     -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>
 {
-    auto *mySettings = MySettings::globalInstance();
-
     ModelInfo modelInfo = ModelList::globalInstance()->defaultModelInfo();
     const QList<ModelInfo> modelList = ModelList::globalInstance()->selectableModelList();
     for (const ModelInfo &info : modelList) {
@@ -779,17 +781,20 @@ auto Server::handleChatRequest(const ChatRequest &request)
     }
     auto startOffset = m_chatModel->appendResponseWithHistory(messages);
 
-    // FIXME(jared): taking parameters from the UI inhibits reproducibility of results
-    backend::GenerationParams genParams {
-        .n_predict      = request.max_tokens,
-        .top_k          = mySettings->modelTopK(modelInfo),
-        .top_p          = request.top_p,
-        .min_p          = request.min_p,
-        .temp           = request.temperature,
-        .n_batch        = mySettings->modelPromptBatchSize(modelInfo),
-        .repeat_penalty = float(mySettings->modelRepeatPenalty(modelInfo)),
-        .repeat_last_n  = mySettings->modelRepeatPenaltyTokens(modelInfo),
-    };
+    std::unique_ptr<GenerationParams> genParams;
+    {
+        using enum GenerationParam;
+        QMap<GenerationParam, QVariant> values;
+        if (auto v = request.max_tokens ) values.insert(NPredict,    *v);
+        if (auto v = request.temperature) values.insert(Temperature, *v);
+        if (auto v = request.top_p      ) values.insert(TopP,        *v);
+        if (auto v = request.min_p      ) values.insert(MinP,        *v);
+        try {
+            genParams.reset(modelDescription()->makeGenerationParams(values));
+        } catch (const std::exception &e) {
+            throw InvalidRequestError(e.what());
+        }
+    }
 
     int promptTokens   = 0;
     int responseTokens = 0;
@@ -797,7 +802,7 @@ auto Server::handleChatRequest(const ChatRequest &request)
     for (int i = 0; i < request.n; ++i) {
         ChatPromptResult result;
         try {
-            result = promptInternalChat(m_collections, genParams, startOffset);
+            result = promptInternalChat(m_collections, *genParams, startOffset);
         } catch (const std::exception &e) {
             m_chatModel->setResponseValue(e.what());
             m_chatModel->setError();
diff --git a/gpt4all-chat/src/store_base.cpp b/gpt4all-chat/src/store_base.cpp
new file mode 100644
index 00000000..f3d49eb6
--- /dev/null
+++ b/gpt4all-chat/src/store_base.cpp
@@ -0,0 +1,164 @@
+#include "store_base.h"
+
+#include <fmt/format.h>
+#include <gpt4all-backend/formatters.h> // IWYU pragma: keep
+
+#include <QByteArray>
+#include <QDebug>
+#include <QIODevice>
+#include <QLatin1StringView> // IWYU pragma: keep
+#include <QSaveFile>
+#include <QUrl>
+#include <QtAssert>
+#include <QtLogging>
+
+#include <array>
+#include <stdexcept>
+#include <string>
+#include <system_error>
+
+namespace fs   = std::filesystem;
+namespace json = boost::json;
+namespace sys  = boost::system;
+using namespace Qt::StringLiterals;
+
+
+namespace gpt4all::ui {
+
+
+DataStoreError::DataStoreError(const QFileDevice *file)
+    : m_error(file->error())
+    , m_errorString(file->errorString())
+{
+    Q_ASSERT(file->error());
+}
+
+DataStoreError::DataStoreError(const boost::system::system_error &e)
+    : m_error(e.code())
+    , m_errorString(QString::fromUtf8(e.what()))
+{
+    Q_ASSERT(e.code());
+}
+
+DataStoreError::DataStoreError(QString e)
+    : m_error()
+    , m_errorString(e)
+    {}
+
+void DataStoreError::raise() const
+{
+    std::visit(Overloaded {
+        [&](QFileDevice::FileError    e) { throw FileError(m_errorString, e); },
+        [&](boost::system::error_code e) { throw std::runtime_error(m_errorString.toUtf8().constData()); },
+        [&](std::monostate             ) { throw std::runtime_error(m_errorString.toUtf8().constData()); },
+    }, m_error);
+    Q_UNREACHABLE();
+}
+
+auto DataStoreBase::reload() -> DataStoreResult<>
+{
+    if (auto res = clear(); !res)
+        return res;
+
+    json::stream_parser parser;
+    QFile file;
+
+    for (auto &entry : fs::directory_iterator(m_path)) {
+        file.setFileName(entry.path());
+        if (!file.open(QFile::ReadOnly)) {
+            qWarning().noquote() << "skipping unopenable file:" << file.fileName();
+            continue;
+        }
+        auto jv = read(file, parser);
+        if (!jv) {
+            (qWarning().nospace() << "skipping " << file.fileName() << "because of read error: ").noquote()
+                << jv.error().errorString();
+        } else if (auto [unique, uuid] = insert(*jv); !unique)
+            qWarning() << "skipping duplicate data store entry:" << uuid;
+        file.close();
+    }
+    return {};
+}
+
+auto DataStoreBase::setPath(fs::path path) -> DataStoreResult<>
+{
+    if (path != m_path) {
+        m_path = std::move(path);
+        return reload();
+    }
+    return {};
+}
+
+auto DataStoreBase::getFilePath(const QString &name) -> std::filesystem::path
+{ return m_path / fmt::format("{}.json", QLatin1StringView(normalizeName(name))); }
+
+auto DataStoreBase::openNew(const QString &name) -> DataStoreResult<std::unique_ptr<QFile>>
+{
+    auto path = getFilePath(name);
+    auto file = std::make_unique<QFile>(path);
+    if (file->exists())
+        return std::unexpected(sys::system_error(std::make_error_code(std::errc::file_exists), path.string()));
+    if (!file->open(QFile::WriteOnly | QFile::NewOnly))
+        return std::unexpected(&*file);
+    return file;
+}
+
+auto DataStoreBase::openExisting(const QString &name) -> DataStoreResult<std::unique_ptr<QSaveFile>>
+{
+    auto path = getFilePath(name);
+    if (!QFile::exists(path))
+        return std::unexpected(sys::system_error(
+            std::make_error_code(std::errc::no_such_file_or_directory), path.string()
+        ));
+    auto file = std::make_unique<QSaveFile>(toQString(path));
+    if (!file->open(QSaveFile::WriteOnly | QSaveFile::ExistingOnly))
+        return std::unexpected(&*file);
+    return file;
+}
+
+auto DataStoreBase::read(QFileDevice &file, boost::json::stream_parser &parser) -> DataStoreResult<boost::json::value>
+{
+    for (;;) {
+        auto chunk = file.read(JSON_BUFSIZ);
+        if (file.error())
+            return std::unexpected(&file);
+        if (chunk.isEmpty()) {
+            Q_ASSERT(file.atEnd());
+            break;
+        }
+        parser.write(chunk.data(), chunk.size());
+    }
+    return parser.release();
+}
+
+auto DataStoreBase::write(const json::value &value, QFileDevice &file) -> DataStoreResult<>
+{
+    m_serializer.reset(&value);
+    std::array<char, JSON_BUFSIZ> buf;
+    while (!m_serializer.done()) {
+        auto chunk = m_serializer.read(buf.data(), buf.size());
+        qint64 nWritten = file.write(chunk.data(), chunk.size());
+        if (nWritten < 0)
+            return std::unexpected(&file);
+        Q_ASSERT(nWritten == chunk.size());
+    }
+
+    if (!file.flush())
+        return std::unexpected(&file);
+
+    return {};
+}
+
+QByteArray DataStoreBase::normalizeName(const QString &name)
+{
+    auto lower = name.toLower();
+    auto norm = QUrl::toPercentEncoding(lower, /*exclude*/ " !#$%&'()+,;=@[]^`{}"_ba, /*include*/ "~"_ba);
+
+    // "." and ".." are special filenames
+    return norm == "."_ba  ? "%2E"_ba    :
+           norm == ".."_ba ? "%2E%2E"_ba :
+           norm;
+}
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/store_base.h b/gpt4all-chat/src/store_base.h
new file mode 100644
index 00000000..79ba2300
--- /dev/null
+++ b/gpt4all-chat/src/store_base.h
@@ -0,0 +1,119 @@
+#pragma once
+
+#include "utils.h" // IWYU pragma: keep
+
+#include <boost/json.hpp> // IWYU pragma: keep
+#include <boost/system.hpp> // IWYU pragma: keep
+#include <tl/generator.hpp>
+
+#include <QFile>
+#include <QFileDevice>
+#include <QString>
+#include <QUuid>
+#include <QtTypes> // IWYU pragma: keep
+
+#include <expected>
+#include <filesystem>
+#include <memory>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <variant>
+
+class QByteArray;
+class QSaveFile;
+
+
+namespace gpt4all::ui {
+
+
+class DataStoreError {
+public:
+    using ErrorCode = std::variant<
+        QFileDevice::FileError,
+        boost::system::error_code,
+        std::monostate
+    >;
+
+    DataStoreError(const QFileDevice *file);
+    DataStoreError(const boost::system::system_error &e);
+    DataStoreError(QString e);
+
+    [[nodiscard]] const ErrorCode &error      () const { return m_error;       }
+    [[nodiscard]] const QString   &errorString() const { return m_errorString; }
+
+    [[noreturn]] void raise() const;
+
+private:
+    ErrorCode m_error;
+    QString   m_errorString;
+};
+
+template <typename T = void>
+using DataStoreResult = std::expected<T, DataStoreError>;
+
+class DataStoreBase {
+protected:
+    explicit DataStoreBase(std::filesystem::path path)
+        : m_path(std::move(path))
+        {}
+
+public:
+    auto path() const -> const std::filesystem::path & { return m_path; }
+    auto setPath(std::filesystem::path path) -> DataStoreResult<>;
+
+protected:
+    auto reload() -> DataStoreResult<>;
+    virtual auto clear() -> DataStoreResult<> = 0;
+    struct InsertResult { bool unique; QUuid uuid; };
+    virtual InsertResult insert(const boost::json::value &jv) = 0;
+
+    // helpers
+    auto getFilePath(const QString &name) -> std::filesystem::path;
+    auto openNew(const QString &name) -> DataStoreResult<std::unique_ptr<QFile>>;
+    auto openExisting(const QString &name) -> DataStoreResult<std::unique_ptr<QSaveFile>>;
+    static auto read(QFileDevice &file, boost::json::stream_parser &parser) -> DataStoreResult<boost::json::value>;
+    auto write(const boost::json::value &value, QFileDevice &file) -> DataStoreResult<>;
+
+private:
+    static constexpr uint JSON_BUFSIZ = 16384; // default QFILE_WRITEBUFFER_SIZE
+
+    static QByteArray normalizeName(const QString &name);
+
+protected:
+    std::filesystem::path m_path;
+
+private:
+    boost::json::serializer m_serializer;
+};
+
+template <typename T>
+class DataStore : public DataStoreBase {
+public:
+    explicit DataStore(std::filesystem::path path);
+
+    auto list() -> tl::generator<const T &>;
+    auto setData(T data) -> DataStoreResult<>;
+    auto remove(const QUuid &id) -> DataStoreResult<>;
+
+    auto acquire(QUuid        id) -> DataStoreResult<const T *>;
+    auto release(const QUuid &id) -> DataStoreResult<>;
+
+    [[nodiscard]]
+    auto operator[](const QUuid &id) const -> const T &
+    { return m_entries.at(id); }
+
+protected:
+    auto createImpl(T data, const QString &name) -> DataStoreResult<const T *>;
+    auto clear() -> DataStoreResult<> final;
+    InsertResult insert(const boost::json::value &jv) override;
+
+private:
+    std::unordered_map<QUuid, T> m_entries;
+    std::unordered_set<QUuid>    m_acquired;
+};
+
+
+} // namespace gpt4all::ui
+
+#include "store_base.inl" // IWYU pragma: export
diff --git a/gpt4all-chat/src/store_base.inl b/gpt4all-chat/src/store_base.inl
new file mode 100644
index 00000000..840afd0d
--- /dev/null
+++ b/gpt4all-chat/src/store_base.inl
@@ -0,0 +1,128 @@
+#include "json-helpers.h" // IWYU pragma: keep
+
+#include <boost/json.hpp> // IWYU pragma: keep
+#include <gpt4all-backend/json-helpers.h> // IWYU pragma: keep
+
+#include <QSaveFile>
+#include <QtAssert>
+
+
+namespace gpt4all::ui {
+
+
+template <typename T>
+DataStore<T>::DataStore(std::filesystem::path path)
+    : DataStoreBase(std::move(path))
+{
+    if (auto res = reload(); !res)
+        res.error().raise(); // should be impossible
+}
+
+template <typename T>
+auto DataStore<T>::list() -> tl::generator<const T &>
+{
+    for (auto &[_, value] : m_entries)
+        co_yield value;
+}
+
+template <typename T>
+auto DataStore<T>::createImpl(T data, const QString &name) -> DataStoreResult<const T *>
+{
+    // acquire path
+    auto file = openNew(name);
+    if (!file)
+        return std::unexpected(file.error());
+
+    // serialize
+    if (auto res = write(boost::json::value_from(data), **file); !res)
+        return std::unexpected(res.error());
+
+    // insert
+    auto [it, unique] = m_entries.emplace(data.id, std::move(data));
+    Q_ASSERT(unique);
+
+    // acquire data ownership
+    if (auto res = acquire(data.id); !res)
+        return std::unexpected(res.error());
+
+    return &it->second;
+}
+
+template <typename T>
+auto DataStore<T>::setData(T data) -> DataStoreResult<>
+{
+    // acquire path
+    auto file = openExisting(data.name);
+    if (!file)
+        return std::unexpected(file.error());
+
+    // serialize
+    if (auto res = write(boost::json::value_from(data), **file); !res)
+        return std::unexpected(res.error());
+    if (!(*file)->commit())
+        return std::unexpected(file->get());
+
+    // update
+    m_entries.at(data.id) = std::move(data);
+    return {};
+}
+
+template <typename T>
+auto DataStore<T>::remove(const QUuid &id) -> DataStoreResult<>
+{
+    // acquire UUID
+    auto it = m_entries.find(id);
+    if (it == m_entries.end())
+        return std::unexpected(QStringLiteral("id not found: %1").arg(id.toString()));
+
+    auto &[_, data] = *it;
+
+    // remove the path
+    auto path = getFilePath(data.name);
+    QFile file(path);
+    if (!file.remove())
+        throw std::unexpected(&file);
+
+    // update cache
+    m_entries.erase(it);
+    return {};
+}
+
+template <typename T>
+auto DataStore<T>::acquire(QUuid id) -> DataStoreResult<const T *>
+{
+    auto [it, unique] = m_acquired.insert(std::move(id));
+    if (!unique)
+        return std::unexpected(QStringLiteral("id already acquired: %1").arg(id.toString()));
+    return &(*this)[*it];
+}
+
+template <typename T>
+auto DataStore<T>::release(const QUuid &id) -> DataStoreResult<>
+{
+    if (!m_acquired.erase(id))
+        return std::unexpected(QStringLiteral("id not acquired: %1").arg(id.toString()));
+    return {};
+}
+
+template <typename T>
+auto DataStore<T>::clear() -> DataStoreResult<>
+{
+    if (!m_acquired.empty())
+        return std::unexpected(QStringLiteral("cannot clear data store with living references"));
+    m_entries.clear();
+    return {};
+}
+
+template <typename T>
+auto DataStore<T>::insert(const boost::json::value &jv) -> InsertResult
+{
+    auto data = boost::json::value_to<T>(jv);
+    auto id = data.id;
+    auto [_, ok] = m_entries.emplace(id, std::move(data));
+    return { ok, std::move(id) };
+}
+
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/store_provider.cpp b/gpt4all-chat/src/store_provider.cpp
new file mode 100644
index 00000000..e7dc1eec
--- /dev/null
+++ b/gpt4all-chat/src/store_provider.cpp
@@ -0,0 +1,25 @@
+#include "store_provider.h"
+
+#include <utility>
+
+
+namespace gpt4all::ui {
+
+
+auto ProviderStore::create(QString name, QUrl base_url, QString api_key)
+    -> DataStoreResult<const ModelProviderData *>
+{
+    ModelProviderData data { QUuid::createUuid(), ProviderType::openai, name, std::move(base_url),
+                             OpenaiProviderDetails { std::move(api_key) }                          };
+    return createImpl(std::move(data), name);
+}
+
+auto ProviderStore::create(QString name, QUrl base_url)
+    -> DataStoreResult<const ModelProviderData *>
+{
+    ModelProviderData data { QUuid::createUuid(), ProviderType::ollama, name, std::move(base_url) };
+    return createImpl(std::move(data), name);
+}
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/store_provider.h b/gpt4all-chat/src/store_provider.h
new file mode 100644
index 00000000..02666a10
--- /dev/null
+++ b/gpt4all-chat/src/store_provider.h
@@ -0,0 +1,49 @@
+#pragma once
+
+#include "store_base.h"
+
+#include <boost/describe/class.hpp>
+#include <boost/describe/enum.hpp>
+
+#include <QString>
+#include <QUrl>
+#include <QUuid>
+
+#include <variant>
+
+
+namespace gpt4all::ui {
+
+
+BOOST_DEFINE_ENUM_CLASS(ProviderType, openai, ollama)
+
+struct OpenaiProviderDetails {
+    QString api_key;
+};
+BOOST_DESCRIBE_STRUCT(OpenaiProviderDetails, (), (api_key))
+
+struct ModelProviderData {
+    using Details = std::variant<std::monostate, OpenaiProviderDetails>;
+    QUuid        id;
+    ProviderType type;
+    QString      name;
+    QUrl         base_url;
+    Details      details;
+};
+BOOST_DESCRIBE_STRUCT(ModelProviderData, (), (id, type, name, base_url, details))
+
+class ProviderStore : public DataStore<ModelProviderData> {
+private:
+    using Super = DataStore<ModelProviderData>;
+
+public:
+    using Super::Super;
+
+    /// OpenAI
+    auto create(QString name, QUrl base_url, QString api_key) -> DataStoreResult<const ModelProviderData *>;
+    /// Ollama
+    auto create(QString name, QUrl base_url) -> DataStoreResult<const ModelProviderData *>;
+};
+
+
+} // namespace gpt4all::ui
diff --git a/gpt4all-chat/src/utils.h b/gpt4all-chat/src/utils.h
index 5c21574f..338b76c4 100644
--- a/gpt4all-chat/src/utils.h
+++ b/gpt4all-chat/src/utils.h
@@ -1,16 +1,56 @@
 #pragma once
 
+#include <QFileDevice>
+#include <QHash>
 #include <QJsonValue>
+#include <QBitArray> // for qHash overload // IWYU pragma: keep
 #include <QLatin1StringView> // IWYU pragma: keep
 
+#include <concepts>
+#include <filesystem>
+#include <functional>
 #include <initializer_list>
+#include <stdexcept>
 #include <utility> // IWYU pragma: keep
 
 // IWYU pragma: no_forward_declare QJsonValue
 class QJsonObject;
+class QVariant;
+template <typename Key, typename T> class QMap;
 
 
 // alternative to QJsonObject's initializer_list constructor that accepts Latin-1 strings
 QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringView, QJsonValue>> args);
 
+QJsonObject &extend(QJsonObject &obj, const QMap<QLatin1StringView, QVariant> &values);
+
+QString toQString(const std::filesystem::path &path);
+auto    toFSPath (const QString &str) -> std::filesystem::path;
+
+template <typename T>
+concept QHashable = requires(const T &x) {
+    { qHash(x) } -> std::same_as<size_t>;
+};
+
+template <QHashable T>
+struct std::hash<T> {
+    size_t operator()(const T &value) const noexcept
+    { return qHash(value); }
+};
+
+class FileError : public std::runtime_error {
+public:
+    explicit FileError(const QFileDevice *file)
+        : FileError(file->errorString(), file->error()) {}
+    explicit FileError(const QString &str, QFileDevice::FileError code);
+    QFileDevice::FileError code() const noexcept { return m_code; }
+
+private:
+    QFileDevice::FileError m_code;
+};
+
+template <typename... Ts>
+struct Overloaded : Ts... { using Ts::operator()...; };
+
+
 #include "utils.inl" // IWYU pragma: export
diff --git a/gpt4all-chat/src/utils.inl b/gpt4all-chat/src/utils.inl
index 3b2a21a4..68d4b2da 100644
--- a/gpt4all-chat/src/utils.inl
+++ b/gpt4all-chat/src/utils.inl
@@ -1,4 +1,7 @@
 #include <QJsonObject>
+#include <QMap>
+#include <QVariant>
+#include <QtAssert>
 
 
 inline QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringView, QJsonValue>> args)
@@ -8,3 +11,34 @@ inline QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringV
         obj.insert(arg.first, arg.second);
     return obj;
 }
+
+inline QJsonObject &extend(QJsonObject &obj, const QMap<QLatin1StringView, QVariant> &values)
+{
+    for (auto [key, value] : values.asKeyValueRange())
+        obj.insert(key, QJsonValue::fromVariant(value));
+    return obj;
+}
+
+// copied from qfile.h
+inline QString toQString(const std::filesystem::path &path)
+{
+#ifdef Q_OS_WIN
+    return QString::fromStdWString(path.native());
+#else
+    return QString::fromStdString(path.native());
+#endif
+}
+
+// copied from qfile.h
+inline auto toFSPath(const QString &str) -> std::filesystem::path
+{
+    return { reinterpret_cast<const char16_t *>(str.cbegin()),
+             reinterpret_cast<const char16_t *>(str.cend  ())  };
+}
+
+FileError::FileError(const QString &str, QFileDevice::FileError code)
+    : std::runtime_error(str.toUtf8().constData())
+    , m_code(code)
+{
+    Q_ASSERT(code);
+}