Last updated:
Mistral
The French AI lab's model lineup — frontier flagships alongside some of the most-used open models in the industry.
Model Facts
Latest model: Mistral Large 4
- Provider
- Mistral AI
- Released
- Oct 6, 2026
- Context window
- 1M tokens (per Mistral)
- Weights
- Closed
Benchmarks
61.7%
DeepSWE v1.1
Large 4
28.3%
Terminal-Bench 4.0
Large 4
Full Model List
| Model | Released | Status | Note |
|---|---|---|---|
| Mistral Large 4 | Oct 6, 2026 | – | 1.05T-parameter multimodal MoE, public preview |
| Mistral Medium 3.5 | Apr 29, 2026 | – | enterprise reasoning, coding and agentic tasks |
| Mistral Small 4 | Mar 16, 2026 | – | 119B-parameter MoE (~6B active) merging reasoning, vision and coding |
| Mistral 3 family | Dec 1, 2025 | – | Large 3 plus Ministral 14B / 8B / 3B |
| Devstral | May 21, 2025 | legacy | agentic software-engineering model |
| Mistral Large 2 | Jul 24, 2024 | – | prior closed-weight flagship |
| Codestral | May 29, 2024 | legacy | code-specialized model, 80+ programming languages |
| Mixtral 8x7B | Dec 11, 2023 | – | early Mixture-of-Experts open release |
| Mistral 7B | Sep 27, 2023 | – | the lab's first widely-adopted open-weight release |
- Official model page:
- mistral.ai
- Official documentation:
- docs.mistral.ai
- Other sources:
- Mistral releases (BenchLM)
- Mistral Large 4 (MarkTechPost)
What it is
Mistral AI is the French lab behind a mix of open-weight and API models. Its newest flagship is Mistral Large 4 ("Le Chonk"), released as a public preview on October 6, 2026: a multimodal mixture-of-experts model with about 1.05T total parameters (roughly 49–52B active per token — the launch materials changed between the two figures).
Open-weight status
Mistral labels Large 4 open-weight, but as of mid-October 2026 the weights are not yet downloadable; Mistral has promised them by the end of October 2026 and has not named a license. This entry is marked closed-weight until they ship. Earlier releases such as Mistral 7B and the Ministral models are open; Medium 3.5 is available through the API.
Where these numbers come from
The benchmarks are figures Mistral reported at launch: 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.0. The 1M-token context window is Mistral's claim; Artificial Analysis lists 524,288 tokens.
Notable traits
- Large 4 is a trillion-parameter-class model from a European lab.
- Smaller models (Small 4, Ministral) remain popular for self-hosting.
- Voxtral voice models extend the lineup beyond text.