Magistral Review: Mistral's First Reasoning Model
Mistral released Magistral on June 10, 2025 — their first reasoning model. Magistral Medium hits 73.6% on AIME2024. Magistral Small is 24B, Apache 2.0, and already on Hugging Face. Here's what it does, what the benchmarks show, and who should use it.
Reviewed 2026-06-14 · Source: Mistral announcement
Quick verdict
Magistral is a strong reasoning model debut for Mistral. The open-source Small variant (24B, Apache 2.0) is the standout: competitive AIME2024 scores at a parameter count that can run on a single high-end server. The Medium variant's 10x throughput advantage in Le Chat and native multilingual chain-of-thought give it clear differentiation from o3-mini and DeepSeek-R1 for enterprise and multilingual use cases. The main limitation: Magistral Medium is closed-weights and Preview-only at launch, making production commitments uncertain until GA.
Benchmarks
| Benchmark | Medium | Small (24B) | Note |
|---|---|---|---|
| AIME 2024 (pass@1) | 73.6% | 70.7% | Math olympiad benchmark; o3-mini reaches ~80% for reference |
| AIME 2024 (majority @64) | 90% | 83.3% | Consensus voting across 64 samples — practical upper bound |
| Le Chat throughput | 10x vs. competitors | — | Flash Answers mode; measured against ChatGPT in Mistral benchmarks |
| Parameters | Undisclosed (enterprise) | 24B | Magistral Small weights on Hugging Face |
| License | Proprietary | Apache 2.0 | Small: commercial use, fine-tuning, redistribution — no restrictions |
What's new
Chain-of-thought that works in your language
Magistral reasons natively in 8+ languages including Arabic, Russian, and Simplified Chinese — not just English with a translation wrapper. The model's chain-of-thought traces appear in the same language as the prompt, which matters for multilingual enterprises reviewing AI reasoning for compliance or audit purposes.
Transparent, traceable reasoning
Every Magistral response shows its logical steps. Unlike opaque model outputs, you can follow the reasoning path, identify where the model applied domain knowledge, and flag errors before they propagate into decisions. This is the core design requirement for regulated industries: finance, legal, healthcare, and government.
Flash Answers in Le Chat
Magistral Medium runs with up to 10x faster token throughput in Le Chat via Flash Answers mode — faster than ChatGPT in Mistral's head-to-head benchmarks. For use cases where reasoning speed matters (customer-facing analysis, real-time decision support), this is a meaningful differentiator over slower thinking models.
Magistral Small: open-source reasoning at 24B
Magistral Small is a 24B parameter model released under Apache 2.0 — the most permissive license available. You can download the weights from Hugging Face (mistralai/Magistral-Small-2506), fine-tune on proprietary domain data, and deploy on-premises with no usage fees or royalties. At launch, 70.7% on AIME2024 made it the strongest open-source reasoning model at the 24B scale.
Think mode for deep reasoning sessions
Beyond Flash Answers, Le Chat's Think mode activates extended chain-of-thought for problems that benefit from longer deliberation — similar to GPT-o1's extended thinking, but with Mistral's multilingual advantage intact. Useful for complex legal research, multi-constraint optimization, or strategic planning tasks.
Pricing
Apache 2.0. Download from Hugging Face: mistralai/Magistral-Small-2506. Self-host on vLLM, llama.cpp, or Transformers. GPU requirements vary by deployment target; 24B weights require ~48GB VRAM for full precision.
Available via La Plateforme API. Check mistral.ai/pricing for current per-token rates under the Magistral or reasoning model tier. Preview access was available at launch with production rates published post-preview.
Flash Answers and Think mode available in Le Chat. Free tier has usage limits. Pro/Team unlocks priority access and extended sessions. Magistral Medium powers the reasoning features in Le Chat.
Magistral Medium available via SageMaker for enterprise teams who need managed infrastructure without running their own Mistral API stack. IBM WatsonX, Azure AI, and Google Cloud Marketplace access announced for future availability.
FAQ
What is Magistral?
Magistral is Mistral AI's first reasoning model, released June 10, 2025. It uses chain-of-thought reasoning to work through complex, multi-step problems — similar to OpenAI's o-series or DeepSeek-R1. It comes in two variants: Magistral Small (24B, Apache 2.0 open-source) and Magistral Medium (enterprise, closed weights).
How does Magistral perform on AIME 2024?
Magistral Medium scores 73.6% on AIME 2024 pass@1, reaching 90% with majority voting across 64 samples. Magistral Small scores 70.7% pass@1 (83.3% @64). For reference, o3-mini high scored around 87% and DeepSeek-R1 scored 72.6% at launch — Magistral Medium is competitive at release.
What languages does Magistral reason in?
Magistral reasons natively in English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. The chain-of-thought appears in the prompt language — not just the final answer. This is a key differentiator for multilingual enterprise use cases where auditability of the reasoning process matters.
Is Magistral Small truly open source?
Yes. Magistral Small is released under Apache 2.0, which is a permissive open-source license. You can use it commercially, fine-tune it on proprietary data, modify the weights, and redistribute without additional restrictions. Weights are available on Hugging Face at mistralai/Magistral-Small-2506.
How do I access Magistral Medium?
Magistral Medium is available via Mistral's La Plateforme API (pay-per-token), in Le Chat (free tier with usage limits, Pro for extended access), and on Amazon SageMaker. IBM WatsonX, Azure AI, and Google Cloud Marketplace access was announced as coming soon at launch.
What is Flash Answers in Le Chat?
Flash Answers is a mode in Le Chat that delivers Magistral Medium responses at up to 10x the token throughput of competing reasoning models. It enables real-time reasoning feedback — Mistral benchmarked it against ChatGPT and showed a significant speed advantage for the same reasoning quality.
What are Magistral's best use cases?
Magistral is designed for tasks requiring multi-step reasoning with transparency: legal research (traceable conclusions), financial modeling (auditable calculations), healthcare decision support (compliance-ready reasoning), software architecture planning, and structured creative writing. It's not optimized for pure conversational tasks — for those, Mistral Small or Medium non-reasoning models are faster and cheaper.
Try Magistral
Magistral Small weights are on Hugging Face under Apache 2.0. Magistral Medium is available via La Plateforme API and Le Chat.
Related models
Affiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.