Magistral

model
Mistral

Reasoning model family trained with pure reinforcement learning — no distillation from external reasoning models. ~50% boost in AIME-24 (pass@1) over the Mistral Medium 3 base checkpoint.

RL on text maintains multimodal understanding, instruction following, and function calling. Magistral Small (24B, Apache 2.0) and Magistral Medium (proprietary).

Paper (arXiv)HuggingFace (Small)OpenRouterArtificial Analysis

Model Details

Architecture DENSE
AA Intelligence 11 was 12

Variants

Name Parameters Notes
Magistral Small 1.2 24B Apache 2.0 open weights; AA Intelligence Index 11, AAOI 50
Magistral Medium 1.2 — Proprietary; AA Intelligence Index 18, AAOI 27.8

Paper

Citations 2
arXiv HTML
reasoningopen-weight