Mistral’s first reasoning model, Magistral, launches with large and small Apache 2.0 version

European IA central Mistral launched the masterful todayA new family of large -language models (LLM) which marks the first of the company to enter the increasingly competitive space of “reasoning” or models that take the time to think about their reflection to catch errors and resolve more complex tasks than LLM based on text.

The announcement includes a double strategic version: a powerful masterful owner for corporate customers and, in particular, an open-source version of 24 billion parameters, masterful Small.

This latest version seems to be calculated to strengthen the company’s commitment to its fundamental roots, after a period when it was faced with criticisms to look at more closed and owners such as its Medium 3 for Enterprises, launched in May 2025.

A return to the open source roots

In a decision that will undoubtedly be celebrated by developers and the wider community of AI, Mistral publishes a masterful little under the license of Apache 2.0 open source permissive.

It is a crucial detail. Unlike more restrictive licenses, Apache 2.0 allows anyone to use freely, modify and distribute the source code of the model, even for commercial purposes.

This allows startups and established companies to create and deploy their own applications in addition to the latest Mistral reasoning architecture without license costs or fear of supplier locking.

This open approach is particularly important given the context. While Mistral has built its reputation on powerful open models, its recent version of Medium 3 as a purely owner has aroused the concern of certain districts of the open source community, which feared that the company will derive towards a more closed ecosystem, similar to competitors like Openai.

The release of a masterful Small under such a permissive license serves as a powerful counterattack, reaffirming the dedication of Mistral to arm the open community with advanced tools.

Competitive performance against formidable enemies

Mistral not only speaks a big game; He came with receipts. The company has published a series of references between the lecturers’ mediums and its own predecessor, Mistral-Medium 3, and its Deepseek competitors. The results show a fiercely competitive model in the reasoning arena.

On the reference in Mathematics Aime-24, masterful-medium marks an impressive 73.6% on precision, neck and neck with its predecessor and considerably surprising the Deepseek models. When using the majority vote (a technique where the model generates several responses and the most common is chosen), its performance on AIM-24 jump at 90% staggering.

The new model also holds the blows through other demanding tests, notably GPQA Diamond, a benchmark of answers to higher level and LiveCodebench questions for the coding of the challenges.

While Deepseek-V3 shows solid performance on certain references, the masterful-medium is constantly a high-level reasoning model, validating Mistral’s assertions on its advanced capacities.

Corporate power

While the little lectures are aimed at the open-source world, the validated reference Masterful medium is rightly for the company.

It is accessible via the Mistral Chat interface and the API the platform, it offers the high level performance necessary for critical mission tasks.

Mistral makes this model available on the main cloud platforms, including Amazon Sagemaker, with Azure AI, IBM Watsonx and Google Cloud Marketplace to follow.

This double release strategy allows Mistral to have its cake and eat it too: promoting a dynamic ecosystem around its open models while monetizing its most powerful and tested technology for businesses.

Cost comparison

Regarding the cost, Mistral positions the masterful support as a distinct premium offer, even compared to its own models.

At $ 2 per million entry tokens and $ 5 per million production tokens, it represents a significant increase in prices compared to the former Mistral Medium 3, which costs only $ 0.40 for the contribution and $ 2 for production.

However, when placed against its external rivals, the pricing strategy of masterful medium seems very aggressive. Its entry cost corresponds to that of the latest OpenAi model and is in the Gemini 2.5 Pro range, but its production price of $ 5 substantially subducing both, which are at $ 8 and more than $ 10, respectively.

Although it is considerably more expensive than specialized models like Deepseek-Reasoner, it is an order of magnitude cheaper than the Claude Opus 4 lighthouse of Anthropic, which makes it a convincing value proposal for customers looking for ultra-modern reasoning without paying the highest market prices.

Reasoning that you can see, understand and use

Mistral pushes three basic advantages with the masterful line: transparency, multilingualism and speed.

Breaking with the nature of the “black box” many AI models, the masterful is designed to produce a traceable “chain of thought”. This allows users to follow the logical path of the model, a critical characteristic for professional fields with high issues such as law, finance and health care, where conclusions must be verifiable.

In addition, these reasoning capacities are global. Mistral emphasizes the “multilingual dexterity” of the model, highlighting high fidelity performance in languages, notably French, Spanish, German, Italian, Arabic, simplified and simplified Chinese.

In terms of performance, the company claims to increase major speed. A new feature “Think mode” and “Flash Answers” in the cat would allow masterful medium to achieve up to 10 times the token flow of competitors, facilitating reasoning in real time on an invisible scale.

From generation of the code to creative strategy and beyond

The requests for masterful are vast. Mistral targets any use of use that requires structured accuracy and thinking, financial modeling and legal analysis to software architecture and data engineering. The company even presented the capacity of the model to generate a physical simulation at a blow, demonstrating its understanding of complex systems.

But it’s not all business. Mistral also recommends the model as a “creative companion” for writing and narration, capable of producing work which is either very coherent, or, as the Society says, “deliciously eccentric”.

With the masterful, Mistral IA makes a strategic game not only to compete, but to lead in the next border of the AI. By re -engaging its open source base with a powerful and authorized model with a license while simultaneously pushing the limits of business quality performance, the company indicates that the future of reasoning IA will be both powerful and, in a significant way, open to all.