Magistral by Mistral AI: Advancing Transparent Reasoning in Enterprise AI

Magistral in short: Magistral is the first reasoning model from Mistral AI, released on June 10, 2025, in two versions: Magistral Small, a 24-billion-parameter model under the Apache 2.0 license, and Magistral Medium, which was available only through Mistral’s API and partner platforms. Both show their reasoning steps in the answer, and according to Mistral, they reason in eight languages, including German.

Today, Magistral is no longer the right choice for new projects. Mistral lists all Magistral versions under “Deprecated & retired” in its model overview and names Mistral Small 4 and Mistral Medium 3.5 as replacements. The weights of Magistral Small remain under Apache 2.0 and can still be run locally.

This guide covers what Magistral does, how to read its benchmarks, how to run it locally, and what to use instead.

Magistral at a Glance

What Is Magistral?

Magistral is the name of Mistral AI’s reasoning model line. A reasoning model does not jump straight to the answer. It first produces a chain of intermediate steps and derives the result from them. According to its model card, Magistral Small is based on Mistral Small 3.1, was fine-tuned on reasoning traces from Magistral Medium, and then refined with reinforcement learning. Mistral describes the method in the paper “Magistral” on arXiv. With version 1.2, Magistral became a multimodal reasoning model according to Mistral’s documentation.

Key Features

Visible reasoning steps

Magistral outputs its intermediate steps, which makes results easier to check for calculations, contract clauses, or code. Visible steps don’t guarantee a correct result, though: a plausible-sounding chain of thought can still be wrong. Expert review remains necessary.

Multilingual reasoning

According to Mistral, Magistral reasons in English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. Mistral makes no equivalent promise for other languages.

Think mode and Flash Answers in Mistral Vibe

At launch, Mistral added a Think mode and “Flash Answers” to Mistral Vibe. With them, Magistral Medium reaches up to ten times higher token throughput than most competitors, according to Mistral. That is a vendor claim from the time of release.

Open weights for Magistral Small

Magistral Small is released under Apache 2.0 and may be used commercially. Magistral Medium was an API-only model.

Benchmarks: What Mistral Reports

The figures come from Mistral at release (Magistral announcement, version 1.0). Math problems say little about everyday writing or professional research. For how other vendors’ reasoning models compare, see our posts on Claude Opus 5 and the GPT-5.6 review.

Running Magistral Locally: GGUF, Ollama and vLLM

According to its model card, Magistral Small runs locally on a single RTX 4090 or a MacBook with 32 GB of RAM once it is quantized. Three common routes:

• Ollama: the easiest start; the model is called “magistral” in the Ollama library.

• llama.cpp and LM Studio: with quantized GGUF files, for example from Unsloth.

• vLLM: recommended by Mistral for production inference.

A quick test with Ollama:

ollama run magistral

Quantization saves memory and costs some quality. Magistral Medium isn’t available as open weights. The benefit for companies: with local operation, inputs never leave your own infrastructure. If you plan such a setup, our AI development team can help. For background on the trend, read Are EU companies building AI on-premises?

As an alternative, Mistral Small 4 is available as open weights under Apache 2.0. It has 119 billion parameters (6.5 billion active), so it needs considerably more memory than a 24B model.

Magistral Compared With Its Successors

Prices in US dollars per Mistral’s documentation. They change, so check the model pages before booking. For the whole model family, see our guide What is Mistral AI?. For the earlier model line, read about Mistral Medium 3.

How Much Does Magistral Cost, and Is It Free?

The Magistral Small and Medium API models can no longer be booked. For a Mistral reasoning model through the API, Mistral Small 4 costs $0.15 per million input tokens and $0.60 per million output tokens, and Mistral Medium 3.5 costs $1.50 and $7.50. The weights of Magistral Small are free to use, so only hardware and power costs apply. In the chat product Mistral Vibe, Mistral offers a free plan with image generation, 500 memories, and more than 100+ connectors. Paid plans and current prices are on mistral.ai/pricing.

Data Protection and GDPR

Mistral AI is based in France, so the GDPR applies. Whether your use is compliant depends on your configuration, your contracts (for example, data processing agreements), and the data you process. When you run Magistral Small locally, data doesn’t leave your infrastructure, but your data protection officer should still review the setup. More on the topic: European data sovereignty and the overview: What is Mistral AI?. This is not legal advice.

Who Should Use Magistral Today?

• Existing projects on the Magistral API: Mistral no longer provides the models. Switch calls to Mistral Small 4 or Mistral Medium 3.5 and keep model names in a config file instead of in code.

• Local operation and research: Magistral Small (24B, Apache 2.0) stays interesting if a small open reasoning model should run on your own hardware.

• New projects: start with Mistral Small 4 or Medium 3.5 and test against your own data.

• Benchmarking: Magistral remains a reference to measure how much its successors improved.

At launch, Mistral named legal research, financial forecasting, software development, and regulated industries as target areas. Typical examples are checking contract clauses, tracing calculations, and planning code refactors. No language model is suitable as the sole basis for medical, legal, or financial decisions.

Conclusion

Magistral was Mistral’s first step into reasoning models: visible reasoning steps, eight languages, and an open 24B model. Today it suits local operation and comparisons, while new projects are better started on its successors. If you want help choosing the right model for your budget, data protection needs, and tasks, talk to our AI consulting team. For what to look for in an advisor, see how to choose an AI consultant.

FAQ

What is Magistral?

Magistral is the first reasoning model from Mistral AI. It shows its reasoning steps and was released on June 10, 2025, as Magistral Small (24B, open weights) and Magistral Medium (API only).

When was Magistral released?

On June 10, 2025. Version 1.1 followed in July 2025, and version 1.2 in September 2025.

Is Magistral open source and free?

Magistral Small has open weights under Apache 2.0 and may be used commercially. Only hardware and power costs apply. Magistral Medium was API-only and is no longer available.

Can I run Magistral locally (GGUF, vLLM, Ollama)?

Yes. According to its model card, quantized Magistral Small runs on an RTX 4090 or a MacBook with 32 GB of RAM. Ollama is the easiest route, and Mistral recommends vLLM for production.

Is Magistral still current?

No. Mistral lists all Magistral versions under “Deprecated & retired” and names Mistral Small 4 and Mistral Medium 3.5 as replacements.

How is Mistral AI different from ChatGPT?

Mistral AI releases many models with open weights and is based in the EU. ChatGPT is a product from OpenAI, and its main models run on OpenAI’s platform. Mistral’s chat product is called Mistral Vibe. Which fits better depends on your data, team and budget.

Is Mistral AI GDPR compliant?

Mistral is based in France, and the GDPR applies. Whether your use is compliant depends on configuration, contracts, and data use. With local operation of open models, data stays on your infrastructure.

Can I use Mistral AI for free, including for images?

Yes. According to Mistral, Vibe has a free plan that includes image generation. Usage is limited, and the paid plans have higher limits. Current details are on mistral.ai/pricing.

Sources

• Mistral: Magistral announcement

• Mistral Docs: Models overview

• arXiv: Magistral (2506.10909)

• Ollama: magistral.