Magistral 1.0 (Small & Medium)

// Overview

Magistral 1.0 is Mistral AI's first family of reasoning models, announced on June 10, 2025. It ships in two sizes: Magistral Small, a 24-billion-parameter open-weight model released under the Apache 2.0 license (model ID magistral-small-2506), and Magistral Medium, a larger, more capable API-only model (magistral-medium-2506) available in preview on Le Chat and Mistral's La Plateforme API. Both are designed to 'think things through' — producing an explicit, verifiable reasoning trace rather than answering in a single step.

Magistral Small is built from Mistral Small 3.1 by adding reasoning through supervised fine-tuning on traces distilled from Magistral Medium, followed by reinforcement learning. Because it is 24B and Apache-2.0 licensed, it can be self-hosted and, in quantized form, run on a single high-memory consumer GPU. Magistral Medium is built on Mistral Medium 3 and trained with a from-scratch reinforcement-learning pipeline; Mistral highlights it for speed, claiming up to roughly 10x faster token throughput than many competitors when used through Le Chat's 'Flash Answers'.

At launch both 1.0 models are text-only and emphasize domain-specific, transparent, and multilingual reasoning, with strong support for languages including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. They are positioned for structured, multi-step work — calculations, programmatic logic, decision trees, and rule-based systems — where a traceable thought process matters. On math and science benchmarks the 1.0 models trail the strongest closed reasoning models of the period (e.g. Gemini 2.5 Pro, Claude Opus 4), but Magistral Small was notable as one of the few openly licensed reasoning models of its time.

// Benchmarks

1. AIME 2024 (pass@1)73.6%
2. AIME 2025 (pass@1)64.9%
3. GPQA Diamond70.8%
4. LiveCodeBench v559.4%
5. AIME 2024 (pass@1)70.7%
6. AIME 2025 (pass@1)62.8%
7. GPQA Diamond68.2%
8. LiveCodeBench v555.8%

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

// Pricing

Mistral La Plateforme API pricing. Reasoning models emit verbose chain-of-thought, so output-token usage (and cost) per request tends to be higher than non-reasoning models. Magistral Small weights are also free to self-host under Apache 2.0.

Pricing source ↗

// Strengths

• Magistral Small ships under Apache 2.0 with downloadable weights — fully self-hostable and commercially usable, rare for a reasoning model at launch
• Transparent, traceable chain-of-thought reasoning you can read and verify, rather than a single opaque answer
• Strong multilingual reasoning across 20+ languages including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese
• Magistral Medium emphasizes throughput — Mistral claims up to ~10x faster token speed than many competitors via Le Chat Flash Answers
• Small is compact (24B), so it runs on a single high-memory GPU when quantized, keeping reasoning workloads on-prem
• Available across many endpoints (Mistral La Plateforme, Together AI, OpenRouter, NVIDIA NIM) plus self-deployment

// Best for

• Self-hosted, privacy-sensitive reasoning workloads using the open-weight Magistral Small
• Multi-step math, logic, and structured-calculation tasks that benefit from an explicit reasoning trace
• Multilingual reasoning and analysis across European, Arabic, Russian, and Chinese text
• Auditable enterprise decision support — decision trees, rule-based systems, and programmatic logic where the thought process must be inspectable
• Coding and software-reasoning tasks (evaluated on LiveCodeBench) within agentic or assistant pipelines
• Low-latency assistant experiences via Magistral Medium's high-throughput 'Flash Answers' on Le Chat

// 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.