Mage VL
Mage VL is an open-weights language model from Microsoft, released July 25, 2026. It has 4.7B parameters (dense). The context window is 256K tokens.
Check it against your own PC →
Can you run Mage VL locally?
The smallest common GPU it runs on is 6 GB of VRAM (with 32 GB of RAM): Q6_K, split GPU / CPU, roughly 13 tokens per second. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.
| GPU memory | Verdict | Best quant | VRAM | RAM | tok/s |
|---|---|---|---|---|---|
| 6 GB VRAM | Runs, slowly | Q6_K (≈ lossless) | 5.2 GB | 1.7 GB | ~13 |
| 8 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 6.9 GB | – | ~40 |
| 12 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 6.9 GB | – | ~53 |
| 16 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 6.9 GB | – | ~72 |
| 24 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 6.9 GB | – | ~130 |
| 32 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 6.9 GB | – | ~220 |
Ready-made quantizations on Hugging Face: JohnTdi/Mage-VL-GGUF · SandLogicTechnologies/mage-vl-GGUF · sahilchachra/Mage-VL-AWQ · ajh-code/Mage-VL-XPO3-NVFP4-W4A4 · mlx-community/Mage-VL-8bit · mlx-community/Mage-VL-OptiQ-4bit
Benchmark results
Epoch AI has not published results for Mage VL yet. New models are usually evaluated within days to weeks.
Links
Hugging Face: microsoft/Mage-VL
More from Microsoft
MAI-Cyber-1.1-Flash · MAI-Base-2 · MAI-Cyber-1-Flash · MAI-Code-1.1-Flash · MAI-Thinking-1 · MAI-Code-1-Flash · Phi-4 · phi-3-small 7.4B