Mistral Large 4
Mistral Large 4 is Mistral's October 2026 general model for text and images. Its mixture-of-experts design activates 52 billion parameters from about 1.05 trillion total parameters. A separate 1.6 billion parameter vision encoder handles images. It is available through Mistral's API and as downloadable weights.
Understand the two active parameter counts
The model card reports 49 billion active parameters per token, or 52 billion when embeddings and the output layer are counted. These describe the same checkpoint. Its reasoning effort can be low, medium or high, with high as the default. The release also includes a DSpark draft head for speculative decoding and a separate NVFP4 checkpoint for Blackwell GPUs.Mistral Large 4 official model card
Check how the service will expose the model
The license defines Model-as-a-Service around third-party access with meaningful control over inputs, parameters or training data. It excludes certain embedded end-user features and simple relaying to models hosted by others. These details matter when assessing a hosted product.Mistral Large 4 license
Use a realistic document workload
Practical guidance. A useful comparison should include the complete document set, image resolutions, requested output and tool calls. The active parameter count describes computation per token. It does not mean the entire checkpoint has the storage needs of a dense 52B model. For self-hosting, measure memory and throughput with the chosen precision and context length.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | mistral-large-4Mistral-Large-4-1T-A52BLe Chonk |
| Context window | 1,000,000 tokens The API model page states 1M. This is the advertised context size, not a measured effective recall limit. Mistral Large 4 API model page |
| Release | 6 October 2026; general availabilityMistral Large 4 API model page |
| Inputs and output | Text and images in; text outMistral Large 4 official model card |
| Expert routing | 128 routed experts; four selected per token, plus one shared expertMistral Large 4 official model card |
| Weight license | Mistral Ecosystem Open-Weight License; additional terms for Model-as-a-Service businessesMistral Large 4 license |
| Open-weight release | 29 October 2026; also called Le ChonkMistral Large 4 (Le Chonk) open-weight release |
Related models
- Mistral Small 4
Mistral AI · Mistral Small
An Apache-licensed multimodal model that combines instruction following, reasoning and coding with selectable reasoning modes.
Launch date not confirmedAPI · Downloadable - Ministral 3 14B
Mistral AI · Ministral 3
The largest Ministral 3 model provides 14B text-and-image checkpoints for base training, instruction following and reasoning.
Launch date not confirmedAPI · Downloadable