24B (23,572,403,200; dense)
128K declared; 40K recommended (40,960 configured currently; 32,768 in launch HF config)
Apache 2.0
Mistral AI’s first open-weight reasoning checkpoint: a text-only, 40-layer dense decoder Transformer with grouped-query attention (32 query heads, 8 KV heads), derived from Mistral Small 3.1 Instruct 2503 without the parent’s vision encoder. Trained by supervised distillation of Magistral Medium reasoning traces followed by modified GRPO reinforcement learning on verifiable math and coding tasks; not a pure-RL/no-SFT release. The report describes four SFT epochs with 10% general instruction data, then RL. Produces visible <think> reasoning and a final answer, with multilingual reasoning encouraged to match the user’s language; the card lists dozens of supported languages, not merely bilingual reasoning. Publisher evaluations report AIME 2024/2025 pass@1 of 70.68%/62.76%, GPQA Diamond 68.18%, and LiveCodeBench v5 55.84%; AIME majority voting over 64 samples reaches 83.3%/76.7%, distinct from single-sample accuracy. Report evaluations allow up to 40K generated tokens for AIME/code and 32K for other tasks; the published protocol does not specify external-tool assistance, and scores are not fixed-latency guarantees. Strengths include multi-step mathematical, coding and multilingual reasoning, with local deployment and official same-release GGUF packaging; fitting a single RTX 4090 or 32GB MacBook requires quantization. Although the card declares 128K context, it warns that performance may degrade beyond 40K and recommends a 40K model-length cap. Published Small results also remain limited on text-only Humanity’s Last Exam (6.4%). The hosted Medium/Le Chat Flash Answers claim of up to 10x throughput does not describe local Small inference. Closed Magistral Medium and later 2507/2509 checkpoints are excluded.