Magistral Small 1.0
Magistral Small 1.0 adds reasoning to Mistral Small 3.1. Its training combines supervised examples from Magistral Medium with reinforcement learning. The resulting 24B checkpoint uses Apache 2.0.
Context needs a practical limit
The weight card describes a 128K window but warns that performance can degrade beyond 40K. It recommends a 40K maximum model length, matching the historical API’s published context. The larger architecture limit should not be treated as a quality guarantee.Magistral Small 1.0 official model card
The prompt format is version-specific
The initial checkpoint uses a reasoning prompt with think tags before a final response. Later Magistral revisions change the token format. A local server must use the template supplied with this checkpoint instead of copying the template from a newer Magistral release.Magistral Small 1.0 official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | Magistral-Small-2506magistral-small-2506 |
| Context window | 40,000 tokens Historical API advertises 40k. The weight card describes 128k but recommends a 40k maximum because longer inputs can reduce performance. Magistral Small 1.0 model and lifecycle record |
| Weight license | Apache 2.0Magistral Small 1.0 official model card |
| Access and lifecycle | Hosted API retired on 30 November 2025. Downloadable weights remain listed.Magistral Small 1.0 model and lifecycle recordMagistral Small 1.0 official model card |
| Parameters | 24 billionMagistral Small 1.0 official model card |