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

European EU strength Mistral started todayA new family (LLM) new family (LLM), which allocates more complex tasks or models to think more complex tasks from the main text-based LLMS and solving more complex tasks

There is a strategic binary issue of the announcement: a powerful, proposed highway environment for enterprise customers, especially 24 billion, 24 billion parameters, the high-source version, highway.

The last release is calculated to strengthen the relation of the company’s basic demolition roots, such as a period of criticism based on proprietary models Backed in May 3, 2025 for enterprises.

A turn for open source roots

In an action that will be noted by neutral developers and a wider AI community, Mistral, the permitting open source releases the highway under the Apache 2.0 license.

This is an important detail. Unlike more restrictive licenses, Apache 2.0 allows everyone to change and distribute the model source code, even free use for commercial purposes.

This allows the starting and established companies, in the same way, to build and place the applications applications without fear of licensing fees or seller lock on the latest thinking architecture.

This open approach is especially important in view of the context. While the Mistral constructed in strong open models, the last 3 of the incoming proposal was a more closed ecosystem, which is concerned about a quarter of an open source, like an open-source society.

The release of a highway under the license that received such a permit is a powerful counter-narrative, confirms the dedication of the Mistral to arm the open society by the most advanced means.

Competitive performance against giant enemies

The Mistral is not only a great game; Came with receipts. The company left his predecessor, Mistral-Mid-Mid-Middle and the high-middle sect against the opponents of Deepseek. The results show a violent competitive model in the justifiable arena.

AIME-24 is effectively effective with accuracy, neck and neck and neck, neck and neck, neck and neck and neck and neck with neck and neck. When using the majority voting (a technique of one of the most common ones and the most common one of the model), the performance on AIME-24 leaks 90%.

The new model also plays other demanding tests, including Livecodebench and Livecodebench, including GPGA Diamond, Coding Problems, including GPGA Diamond, including other required tests.

Deepseek-V3 proves that some criteria show strong performance, the high-level justification model of the high-level facility, which confirms the claims of the advanced capacity of the Mistral.

Venture power

The trunk is confirmed by a small, open source world, benchmark Highway average The enterprise is directed to the hurricane.

This is the highest performance you need for mission-critical assignments via the Mistral Le Chat interface and La Plateeforme API.

This model, including Azure AI, IBM Watsonx and Google Cloud Marketplace, including Mistral, Amazon Sagemaker, is available on high cloud platforms.

This dual release strategy allows you to have the mistral cake and eat it: the strongest, performance technology for corporate clients support the live ecosystem around open models.

The cost comparison

As for the price, the Mistral main environment determines as a different, premium proposal compared to their models.

In verses of one million dollars and $ 5 million dollars, this represents a significant price increase from the old Mistral Middle 3, which spent only $ 0.40 and $ 2 for entry.

However, when placed against their foreign opponents, the high price strategy of the highway seems highly aggressive. The input value is sits in the range of Openai’s last model and Gemini 2.5, but the cost of $ 5 is significantly over $ 8 and $ 10, respectively, respectively.

Although it is significantly more expensive than models like DeepSseek-Midden, it is cheaper than an anthropy Opus 4, offers an attractive value for customers looking for modern thoughts without paying the highest market prices.

You can understand, understand and use it

The autral highway is pushing three main advantages: transparency, multilingual and speed.

Many AI models are designed to prepare a trunk of “chain of thinking”, which removed from the “black box” nature. This allows users to monitor the logical way of the model, such as high-level professional areas such as law, finance and health needed to be checked.

In addition, this meditation skills are global. Mistral, French, Spanish, German, Italian, Italian, Arabic, Russian and simplified, emphasize that the high-loyal performance of languages ​​includes “multilingual agility”.

The company on the performance front shows the main speed. The feature of a new “Mode” and “Flash Answers” in Le Chat allows you to increase Token to Token to 10 times, facilitating real-time real mode on a previously unprecedented scale of the highway.

From COD GEN to creative strategy and out

Applications for the trunk are wide. Mistral aims to make any use of financial modeling and legal analysis, accurate and structured thinking to the program architecture and information engineering. The company also demonstrated the ability to create a shot physics simulation that demonstrates the understanding of complex systems.

But this is not all the work. The Mistral also recommends that the company is either a high sequence or a “delicious eccentric” or a “creative companion” for the story.

Mistral AI with the trunk, is just a strategic game for not compete, but leads to the next border of AI. At the same time, when the enterprise is forcibly pushed with a licensed model with a licensed model with strong, a licensed model, the company said that all the future of the EU will be in a meaningful way.