Magistral (Mistral)

Magistral is a reasoning-model family released by Mistral AI on June 10, 2025, described by the company as its first reasoning model. It comprises two variants — Magistral Small, distributed with open weights, and Magistral Medium, offered commercially. Mistral positioned the family as its entry into the reasoning-model segment occupied by OpenAI's o1/o3 and DeepSeek R1 (mistral.ai; Source: mistral.ai). The release was accompanied by a technical paper, submitted to arXiv on June 12, 2025 (Magistral (Mistral AI, June 2025)).

Variants

• Magistral Small — a 24-billion-parameter model released with open weights under the Apache 2.0 license and downloadable from Hugging Face (Source: mistral.ai). It was produced by supervised fine-tuning of Mistral Small 3 on reasoning traces from Magistral Medium, followed by reinforcement learning (Magistral (Mistral AI, June 2025)).
• Magistral Medium — the larger and more capable variant, offered commercially. Mistral called it an "enterprise version" (Source: mistral.ai). It was trained from Mistral Medium 3 Instruct by reinforcement learning alone, with no supervised fine-tuning on reasoning traces (Magistral (Mistral AI, June 2025)). Its parameter count has not been disclosed.

The split follows Mistral's pattern of pairing an open-weight model with a commercial one within the same family.

Training and architecture

The technical paper describes a reinforcement-learning stack that the authors say was built on Mistral's own models and infrastructure rather than on traces distilled from earlier reasoning models. The training algorithm is a modified form of Group Relative Policy Optimization; the reward combines checks on formatting, correctness (verified final answers for mathematics, compiled and tested programs for code), a length penalty, and a language-consistency reward intended to keep the chain of thought in the user's language. Training runs on an asynchronous distributed system in which generators run continuously while trainers update weights (Magistral (Mistral AI, June 2025)). The paper reports that reinforcement learning alone yields "a nearly 50% accuracy increase in AIME '24 (pass@1), and 30% on LiveCodeBench (v5)" over the Medium base model, and concludes, contrary to DeepSeek's findings, that reinforcement learning adds to distillation even for a small model (Magistral (Mistral AI, June 2025)). See Reasoning Models and Chain-of-Thought.

Capabilities and benchmarks

Mistral's announcement reported 73.6% for Magistral Medium on AIME 2024 (90% with majority voting over 64 samples) and 70.7% and 83.3% respectively for Magistral Small (Source: mistral.ai). The paper also reports results on AIME 2025, MATH-500, GPQA, LiveCodeBench v5 and v6, Aider Polyglot and the text-only subset of Humanity's Last Exam, all developer-reported, with DeepSeek comparison figures quoted from DeepSeek's own paper rather than re-run (Magistral (Mistral AI, June 2025)). Per-setting scores are recorded on AIME (as a model benchmark), MATH and MATH-500, GPQA, LiveCodeBench, Aider Polyglot and Humanity's Last Exam.

Mistral emphasised multilingual reasoning, naming English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese (Source: mistral.ai). On translated versions of AIME 2024 in six languages, the paper reports Magistral Medium scoring 4.3 to 9.9 points below its English result (Magistral (Mistral AI, June 2025)).

Later versions

On September 18, 2025 Mistral released Magistral Small 1.2 and Magistral Medium 1.2, which it described as minor updates to its Magistral 1.1 models (Source: x.com). The 1.2 models added a vision encoder for image inputs and a 128k-token context window; Mistral's model card for Magistral Small 1.2 states it builds on Mistral Small 3.2 and notes that performance "might degrade past 40k" tokens (Source: huggingface.co; Source: docs.mistral.ai). Mistral's documentation lists both 1.2 models as deprecated, naming Mistral Small 4 and Mistral Medium 3.5 as replacements (Source: docs.mistral.ai; Source: docs.mistral.ai).

Safety and evaluations

The Magistral paper contains no safety evaluations (Magistral (Mistral AI, June 2025)).

Availability

At launch Magistral Small was available for self-deployment from Hugging Face, and a preview of Magistral Medium was offered in Mistral's Le Chat assistant and through its La Plateforme API, as well as on Amazon SageMaker, with IBM watsonx, Azure AI and Google Cloud Marketplace announced as forthcoming. Mistral said a "Flash Answers" mode in Le Chat delivered up to 10 times the token throughput of most competitors, a company claim (Source: mistral.ai).

Release and reception

Mistral announced Magistral on June 10, 2025 (mistral.ai). TechCrunch covered the launch under the headline "Mistral releases a pair of AI reasoning models" (2025-06-10).

The release placed Mistral, a France-based developer, among the companies offering reasoning models, a group that also includes OpenAI, Anthropic, DeepSeek, and Moonshot.

Relationships

• developer: Mistral AI
• depends-on: Magistral (Mistral AI, June 2025) (Mistral AI's technical paper, arXiv 2506.10910)
• depends-on: Mistral Large / Mistral Medium / Mixtral — Magistral Medium is trained from Mistral Medium 3, whose baseline scores the paper reports
• related: OpenAI o-series (o1 → o4-mini), DeepSeek-R1, Kimi K2, Reasoning Models and Chain-of-Thought, Open-Weight Frontier Models.

Sources

• Mistral AI / Magistral announcement (2025-06-10) — mistral.ai
• TechCrunch / Mistral releases a pair of AI reasoning models (2025-06-10)
• Magistral (Mistral AI, June 2025) (arXiv 2506.10910, 2025-06-12)
• Hugging Face model card, Magistral Small 1.2 — huggingface.co
• Mistral documentation, Magistral Small 1.2 and Medium 1.2 model pages