Mistral Small 4
Mistral Small 4 puts instruction following, reasoning and coding in one model. It has 119 billion total parameters and activates 6.5 billion per token. It accepts text and images and returns text. The March 2026 release is available through the API and under Apache 2.0 for local use.
One checkpoint replaces separate task families
The model card describes Small 4 as combining the roles of Instruct, Magistral reasoning and Devstral coding. Reasoning effort is set per request: none gives a direct reply, while high enables reasoning. Function calling works with these modes.Mistral Small 4 official model card
Use the published serving variants
Mistral publishes an EAGLE draft head for speculative decoding and an NVFP4 checkpoint. These are serving aids and precision variants of Small 4, rather than separate task families. The hosted release also supports structured output, predicted output and batch requests.Mistral Small 4 official model cardMistral Small 4 API model page
Compare both reasoning modes
Practical guidance. Use the same real tasks in direct and reasoning modes. A field extraction request may need a short direct result, while a difficult code change may benefit from more reasoning. Compare valid outputs, total response time and tool errors. The word Small does not imply that the full 119B checkpoint fits on a small device.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | mistral-small-2603Mistral-Small-4-119B-2603 |
| Context window | 256,000 tokens The model documentation advertises 256k context. Mistral Small 4 API model page |
| Release | 16 March 2026; general availabilityMistral Small 4 API model page |
| Architecture | Mixture of experts with 128 experts and four active expertsMistral Small 4 official model card |
| License | Apache 2.0Mistral Small 4 API model page |
Related models
- Mistral Large 4
Mistral AI · Mistral Large
A sparse multimodal model with a one-million-token context and 52B active parameters, including embeddings and the output layer.
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 - Codestral 2508
Mistral AI · Codestral
A coding model optimized for frequent editor completions and fill-in-the-middle requests.
Launch date API