Publié : 12 December 2025
Actualisé : 8 hours ago
Fiabilité : ✓ Sources vérifiées
Je mets à jour cet article dès que de nouvelles informations sont disponibles.
📋 Table of Contents
Running generative AI models locally is every developer’s dream. Guillaume guides us step by step to install and configure Ministral 3 with an AMD GPU on Windows. Performance and control, finally within reach?
🚀 Required Configuration and First Steps
The goal is clear: to harness the power of a gaming PC to host LLMs (Large Language Models) without breaking the bank on cloud credits. Guillaume used the following configuration:
- Graphics Card: AMD RX 9070 XT
- Processor: AMD Ryzen 7 9700X
- Operating System: Windows
The latest AMD Adrenalin drivers were installed, as well as the AMD HIP SDK, although its usefulness was not immediately apparent. The adventure begins, but not without pitfalls.
❌ The Initial Failure with Ollama
Ollama, with its attractive user interface, seemed like an ideal starting point. However, the software failed to exploit the AMD graphics card, limiting model inference to the CPU. Despite preliminary support announced for AMD GPUs under Windows and Linux in 2024, compatibility remains limited. A disappointment that prompts exploration of other options.
✨ Jan: The Promising Open Source Alternative
Faced with the Ollama impasse, Jan emerges as an attractive open source alternative. Unlike LM Studio, a proprietary option, Jan offers the transparency and flexibility sought. Here’s how to proceed:
- Jan uses llama.cpp as the backend for model inference.
- By default, Jan installs the CPU version of llama.cpp. You must manually install the version compiled to support the Vulkan API.
- Download the latest version of llama.cpp for Windows x64 (Vulkan), for example version b7356.
- Import the backend via the “Install Backend from File” option in Jan. The application takes care of the decompression.
- Select the appropriate Vulkan backend.
This manipulation allows you to take advantage of the power of the AMD GPU.
🌐 Server Mode Deployment
The next step is to make the model accessible remotely, potentially to relatives. To do this, it is crucial to ensure encrypted communications and bypass firewalls. Guillaume uses…
Key Point: Using a Vulkan backend specifically compiled for your GPU is essential for optimizing performance.
🛠️ Summary Table of Key Steps
| Step | Action | Tool |
|---|---|---|
| 1 | Driver Installation | AMD Adrenalin |
| 2 | Test with Ollama | Ollama |
| 3 | Jan Installation | Jan (Open Source) |
| 4 | llama.cpp Download (Vulkan) | llama.cpp |
| 5 | Vulkan Backend Import | Jan |
Note: AMD GPU compatibility with AI tools is constantly evolving. Check regularly for updates and supported versions.
❔ Frequently Asked Questions
Can I use Ollama with my AMD graphics card on Windows to run AI models?
0 Comments