Magistral 1.1 (Small & Medium)

// Overview

Magistral 1.1 is the July 2025 refresh of Mistral AI's first reasoning family, shipping as two tiers: Magistral Small 1.1 (model ID magistral-small-2507), an open-weight 24B-parameter model under the Apache 2.0 license, and Magistral Medium 1.1 (magistral-medium-2507), a more capable proprietary model served only through Mistral's API and Le Chat. Both are text-only reasoning models that expose their step-by-step thinking, which Mistral wraps in explicit [THINK] and [/THINK] tokens.

Magistral Small 1.1 is built on Mistral Small 3.1, with reasoning added by supervised fine-tuning on traces distilled from Magistral Medium and a reinforcement-learning stage on top. It is small enough to run locally — Mistral notes it fits on a single RTX 4090 or a 32GB MacBook once quantized — and carries a 128k context window, though Mistral recommends capping usable length around 40k where quality holds up best. Magistral Medium 1.1 targets the same reasoning tasks at higher accuracy and is accessible via La Plateforme, Le Chat, and partners such as Amazon SageMaker.

The 1.1 update was an incremental quality pass rather than a new architecture: better tone and formatting (cleaner LaTeX and Markdown), shorter answers on easy prompts, and a much lower chance of falling into infinite generation loops. Magistral's signature trait is transparent, multilingual chain-of-thought — it reasons in the user's language across dozens of languages including French, German, Spanish, Arabic, Russian, Chinese, Hindi, and Japanese. Both 1.1 models were retired on 30 November 2025 and replaced by Magistral 1.2.

// Benchmarks

1. AIME 2024 (pass@1) — Magistral Small 1.170.52%
2. AIME 2025 (pass@1) — Magistral Small 1.162.03%
3. GPQA Diamond — Magistral Small 1.165.78%
4. LiveCodeBench v5 — Magistral Small 1.159.17%
5. AIME 2024 (pass@1) — Magistral Medium 1.172.03%
6. AIME 2025 (pass@1) — Magistral Medium 1.160.99%
7. GPQA Diamond — Magistral Medium 1.171.46%
8. LiveCodeBench v5 — Magistral Medium 1.159.35%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Magistral Small is also free to self-host under Apache 2.0. Prices are Mistral's listed API rates.

Pricing source ↗

// Strengths

• Open weights for the Small tier — 24B parameters under Apache 2.0, deployable locally on a single RTX 4090 or a 32GB MacBook when quantized
• Transparent chain-of-thought reasoning that the user can read and verify
• Strong multilingual reasoning that stays in the user's language across dozens of languages
• Explicit [THINK]/[/THINK] reasoning tokens make it easy to separate thinking from the final answer
• 1.1 fixes from 1.0: cleaner LaTeX/Markdown, shorter answers on easy prompts, far fewer infinite-generation loops

// Best for

• Math, science, and logic problems that benefit from explicit multi-step reasoning
• Multilingual reasoning where the chain-of-thought must stay in the user's language
• Self-hosted or on-device reasoning using the open-weight Magistral Small 1.1
• Coding and structured-problem tasks where a transparent, auditable reasoning trace is valuable
• Research and fine-tuning on an Apache 2.0 reasoning base

// How to access

// Magistral — every version

The full lineage of the Magistral line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.