Mistral released the weights of Mistral Large 4 on Hugging Face on 29 October 2026, after a public preview that started on 6 October. It is Mistral's largest model and one of the largest open-weight models available.
What it is
| Mistral Large 4 | |
|---|---|
| Architecture | Mixture of experts with a vision encoder |
| Parameters | About 1 trillion total, about 50 billion active per token |
| Context | 1M tokens |
| Input | Text and images |
| Weights | mistralai/Mistral-Large-4-1T-A52B, FP8 and NVFP4 |
The licence: open, with one big condition
The Mistral Ecosystem Open-Weight License is described as a modified MIT licence. Internal and commercial use are allowed. Offering the model to others as a hosted service needs a separate commercial licence from Mistral. If you want to run it for your own products and team, you're covered; if you want to sell access to it, talk to Mistral first.
The API price
Through Mistral's API it costs about $0.70 per million input tokens, $0.07 for cached input and $2.10 for output, roughly half the preview price. That puts it between the low-cost coding models and the frontier models on price.
Running it yourself
- Hardware: one reference setup serves it with vLLM on eight NVIDIA H200 GPUs. A quantized version is reported to fit on four B200s.
- NVFP4: the separate NVFP4 checkpoint is meant for NVIDIA Blackwell GPUs, where it saves memory and runs faster.