Mistral AI drops Magistral, its first reasoning model

Mistral AI just dropped Magistral, its first reasoning model family built for transparent, multi-step problem solving. It comes in two distinct variants: Magistral Small, an open-weight 24B model, and Magistral Medium, a more powerful enterprise tier.

Released alongside a comprehensive research paper, the family focuses on making complex deliberation traceable, domain-aware, and multilingual.

What it does and how fast it runs

Unlike early reasoning models that often stumble in non-English contexts, Magistral handles chain-of-thought natively across multiple alphabets and languages. It reasons natively in English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese without awkward back-and-forth translation.

On the benchmark front, both models punch well above their weight. In the AIME2024 math competition benchmark, Magistral Medium scored 73.6%, jumping to 90% with majority voting @64. Even the smaller 24B Magistral Small delivered a 70.7% single-run score and 83.3% with majority voting.

Beyond pure math, Mistral showed off one-shot physics simulations handling gravity and collisions, alongside multi-step API and tool execution for systems engineering.

Speed is the other standout factor. When using the new Think mode with Flash Answers in Mistral's Le Chat interface, Magistral Medium delivers token throughput up to 10x faster than most competitors, making real-time interactive reasoning actually practical.

How to get your hands on it

If you want to run it on your own hardware, Magistral Small is available right now under a permissive Apache 2.0 license. The open weights are hosted directly on Hugging Face under mistralai/Magistral-Small-2506, ready for self-deployment, fine-tuning, and commercial use.

For those wanting raw performance without infrastructure overhead, Magistral Medium is accessible in preview on Le Chat and via API on La Plateforme.

It is also available on Amazon SageMaker, with rollouts planned soon for IBM WatsonX, Azure AI, and Google Cloud Marketplace.

Why it matters

Reasoning models have largely been locked behind massive proprietary APIs, making private enterprise deployment both expensive and rigid. By releasing a 24B reasoning model under Apache 2.0, Mistral gives developers a viable, self-hosted alternative for tasks that demand auditability, such as legal analysis, financial forecasting, and healthcare compliance.

We think Mistral's focus on native multilingual logic and sheer inference speed is the right move. Instead of just chasing benchmark bragging rights, they delivered an accessible open-weight workhorse that smaller engineering teams can immediately adopt and build upon.