Mistral Small 3.0
Mistral Small 3.0 provides a dense 24B text model with a 32K context. Its Apache 2.0 release includes a pretrained base checkpoint and an instruction-tuned version.
A local model for responsive applications
The instruction card describes conversational agents, function calls and local processing of sensitive material. It uses Tekken with roughly 131K vocabulary entries and supports JSON output. The advertised single-device examples depend on quantization rather than full-precision memory requirements.Mistral Small 3.0 official model card
What changed from Small 2
The January release increases the parameter count from the earlier 22B model and changes the public weight license to Apache 2.0. Base-2501 is for further adaptation; Instruct-2501 supplies the chat-oriented training. This version is text-only, before Small 3.1’s vision update.Mistral Small 3.0 official model cardMistral Small 3.0 release announcement
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | Mistral-Small-24B-Base-2501Mistral-Small-24B-Instruct-2501mistral-small-2501 |
| Context window | 32,000 tokens The model catalog advertises 32k context for this version. This is not a measured recall limit. Mistral Small 3.0 model and lifecycle record |
| Weight license | Apache 2.0Mistral Small 3.0 official model card |
| Access and lifecycle | Hosted API retired on 30 November 2025. Downloadable weights remain listed.Mistral Small 3.0 model and lifecycle recordMistral Small 3.0 official model card |
| Parameters | 24 billionMistral Small 3.0 official model card |