diff --git a/gpt4all-backend/llamamodel.cpp b/gpt4all-backend/llamamodel.cpp
index 847006c8..2551bc27 100644
--- a/gpt4all-backend/llamamodel.cpp
+++ b/gpt4all-backend/llamamodel.cpp
@@ -301,8 +301,9 @@ bool LLamaModel::initializeGPUDevice(size_t memoryRequired, const std::string& d
 #endif
 }
 
-bool LLamaModel::initializeGPUDevice(const LLModel::GPUDevice &device)
+bool LLamaModel::initializeGPUDevice(const LLModel::GPUDevice &device, std::string *unavail_reason)
 {
+    bool result = false;
 #if defined(GGML_USE_KOMPUTE)
     ggml_vk_device vkDevice;
     vkDevice.index = device.index;
@@ -310,10 +311,16 @@ bool LLamaModel::initializeGPUDevice(const LLModel::GPUDevice &device)
     vkDevice.heapSize = device.heapSize;
     vkDevice.name = device.name;
     vkDevice.vendor = device.vendor;
-    return ggml_vk_init_device(vkDevice);
+    result = ggml_vk_init_device(vkDevice);
+    if (!result && unavail_reason) {
+        *unavail_reason = "failed to init device";
+    }
 #else
-    return false;
+    if (unavail_reason) {
+        *unavail_reason = "built without kompute";
+    }
 #endif
+    return result;
 }
 
 bool LLamaModel::initializeGPUDevice(int device)
diff --git a/gpt4all-backend/llamamodel_impl.h b/gpt4all-backend/llamamodel_impl.h
index 943437ee..d708ddac 100644
--- a/gpt4all-backend/llamamodel_impl.h
+++ b/gpt4all-backend/llamamodel_impl.h
@@ -27,7 +27,7 @@ public:
     int32_t threadCount() const override;
     std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired) override;
     bool initializeGPUDevice(size_t memoryRequired, const std::string& device) override;
-    bool initializeGPUDevice(const GPUDevice &device) override;
+    bool initializeGPUDevice(const GPUDevice &device, std::string *unavail_reason) override;
     bool initializeGPUDevice(int device) override;
     bool hasGPUDevice() override;
     bool usingGPUDevice() override;
diff --git a/gpt4all-backend/llmodel.h b/gpt4all-backend/llmodel.h
index af3e3ff2..ad33c82d 100644
--- a/gpt4all-backend/llmodel.h
+++ b/gpt4all-backend/llmodel.h
@@ -97,7 +97,12 @@ public:
 
     virtual std::vector<GPUDevice> availableGPUDevices(size_t /*memoryRequired*/) { return std::vector<GPUDevice>(); }
     virtual bool initializeGPUDevice(size_t /*memoryRequired*/, const std::string& /*device*/) { return false; }
-    virtual bool initializeGPUDevice(const GPUDevice &/*device*/) { return false; }
+    virtual bool initializeGPUDevice(const GPUDevice &/*device*/, std::string *unavail_reason = nullptr) {
+        if (unavail_reason) {
+            *unavail_reason = "unsupported model type";
+        }
+        return false;
+    }
     virtual bool initializeGPUDevice(int /*device*/) { return false; }
     virtual bool hasGPUDevice() { return false; }
     virtual bool usingGPUDevice() { return false; }
diff --git a/gpt4all-chat/chatllm.cpp b/gpt4all-chat/chatllm.cpp
index c7e1208d..835a58a3 100644
--- a/gpt4all-chat/chatllm.cpp
+++ b/gpt4all-chat/chatllm.cpp
@@ -282,11 +282,14 @@ bool ChatLLM::loadModel(const ModelInfo &modelInfo)
                         }
                     }
 
-                    if (!device || !m_llModelInfo.model->initializeGPUDevice(*device)) {
-                        emit reportFallbackReason("<br>Using CPU: failed to init device");
+                    emit reportFallbackReason(""); // no fallback yet
+                    std::string unavail_reason;
+                    if (!device) {
+                        // GPU not available
+                    } else if (!m_llModelInfo.model->initializeGPUDevice(*device, &unavail_reason)) {
+                        emit reportFallbackReason(QString::fromStdString("<br>Using CPU: " + unavail_reason));
                     } else {
                         actualDevice = QString::fromStdString(device->name);
-                        emit reportFallbackReason(""); // no fallback yet
                     }
                 }
 
@@ -306,6 +309,7 @@ bool ChatLLM::loadModel(const ModelInfo &modelInfo)
                     // We might have had to fallback to CPU after load if the model is not possible to accelerate
                     // for instance if the quantization method is not supported on Vulkan yet
                     emit reportDevice("CPU");
+                    // TODO(cebtenzzre): report somewhere if llamamodel decided the model was not supported
                     emit reportFallbackReason("<br>Using CPU: unsupported quantization type");
                 }