France’s Mistral Releases Large 4 AI Model as Open Weights

France-based artificial intelligence company Mistral has introduced the open weights of its latest flagship model, Mistral Large 4, also known by its nickname, “Le Chonk.” The new model contains approximately 1 trillion parameters and is designed to improve coding, reasoning and cybersecurity tasks.

The release follows an earlier preview on October 6, 2026. After several additional weeks of reinforcement learning, Mistral says the model has achieved significant improvements across multiple performance benchmarks. Its weights are now available through Hugging Face, giving developers and businesses an opportunity to experiment with the technology.

The launch also reflects Europe’s growing interest in developing powerful AI systems that can compete with offerings from American and Chinese companies.

Mistral Large4 AI Shows Improvements in Coding Performance

One of the major highlights of Mistral Large4 AI is its improved performance in software development and coding-related tasks.

According to Mistral, additional reinforcement learning helped the model improve its Terminal-Bench 4.0 score from 28% in the October preview to 41%. This represents a 13-percentage-point increase.

The company also reported that the latest version achieved this improvement while generating 30% fewer output tokens. Lower token consumption can make AI-powered coding workflows more efficient by reducing the amount of generated text required to complete tasks.

On the Artificial Analysis Coding Agent Index, Mistral Large 4 recorded a score of 48%. Mistral says this is approximately comparable to China’s GLM-5.3, while requiring 47% fewer tokens per task on average.

However, closed-source models continue to demonstrate stronger performance in certain evaluations. Anthropic’s Mythos 5.1 reportedly achieved 60.9% on Terminal-Bench, showing that open-weight models still have room to improve.

A Massive Model With a Mixture-of-Experts Architecture

Mistral Large 4 features a mixture-of-experts architecture that allows it to handle demanding workloads without activating every parameter for each request.

Although the model contains approximately 1.05 trillion parameters, it activates around 52 billion parameters for an individual task. This approach helps distribute computational workloads more efficiently.

The architecture is particularly useful for applications involving complex reasoning, programming and technical problem-solving.

The model also provides three reasoning settings:

• Low reasoning: Designed for straightforward requests that require faster responses.
• Medium reasoning: Offers a balance between response speed and deeper analysis.
• High reasoning: Intended for complicated tasks that need more extensive reasoning.

Mistral trained the model using 3,800 Nvidia Grace Blackwell GPUs across its European data centres. This infrastructure highlights the significant computing resources required to develop and improve trillion-parameter AI systems.

Mistral Targets Cybersecurity With Stronger Guardrails

Cybersecurity is another important area behind the Mistral Large 4 release.

Mistral reported a score of 51% on the Artificial Analysis Cyber Index. The company also claims that the model refuses an average of 96.6% of malicious cybersecurity requests, which it describes as its highest measured refusal rate among open models.

Security companies, including Cisco, CrowdStrike and Cloudflare, have expressed support for the release.

However, making powerful AI systems publicly available also creates potential security concerns. Models capable of understanding software vulnerabilities and assisting with technical investigations could potentially be misused.

Mistral acknowledged these risks during the model’s preview period. The company reportedly observed activity in live traffic that suggested the model could improve the ability to conduct real cyberattacks.

As a result, it is continuing to strengthen the model’s safety restrictions and guardrails.

This highlights an important challenge for AI developers: building systems that can assist legitimate cybersecurity professionals without making harmful activities easier.

Open Weights, Licensing and Developer Access

The availability of Mistral Large 4’s weights is a significant part of the announcement.

Developers can access the model through Hugging Face, allowing them to explore its capabilities and build applications using its underlying technology.

The release uses a modified MIT licence. It allows users to access, modify and commercialise the model, although businesses planning to offer it as a hosted model-as-a-service must obtain a separate commercial licence.

Microsoft is also making Mistral Large 4 available through Microsoft Foundry, providing another route for organisations interested in integrating the model into their workflows.

Open-weight availability offers developers greater flexibility than relying entirely on closed AI services. Organisations can explore custom implementations and adapt the model to their requirements, subject to its licensing conditions.

However, open weights do not automatically mean that every business can operate the model without infrastructure costs or additional permissions.

Mistral’s Growing Role in Europe’s AI Industry

The release arrives during an important period for Europe’s artificial intelligence industry.

Mistral has raised approximately $6 billion over three years. Its latest funding round was a €3 billion Series D led by Samsung, placing the company’s valuation above €21 billion.

These investments demonstrate the financial resources being directed towards European AI development.

Mistral’s leadership believes that making model weights publicly available is important for maintaining access to advanced AI technology. The company has also raised concerns about the reliability of depending on external AI providers, particularly when access restrictions can affect businesses and developers.

However, the release has also attracted questions about the wider implications of powerful open-weight technology. Sophos cybersecurity expert Rafe Pilling described the technology as having both beneficial and potentially harmful applications.

This means the discussion surrounding Mistral Large 4 extends beyond competition between countries. It also involves responsible deployment, commercial licensing and cybersecurity protections.

FAQs

1. What is Mistral Large 4 AI?

Mistral Large 4 is a flagship AI model developed by French company Mistral. It contains approximately 1.05 trillion parameters and uses a mixture-of-experts architecture.

2. Where can developers access Mistral Large 4?

Developers can access its open weights through Hugging Face. It is also available through Microsoft Foundry.

3. What makes Mistral Large 4 different from earlier versions?

The model offers improved coding benchmark results, lower token usage, multiple reasoning settings and enhanced cybersecurity safeguards.

4. Is Mistral Large 4 completely free for commercial use?

Its modified MIT licence permits modification and commercial use. However, businesses offering it as a hosted model-as-a-service need a separate commercial licence.

Conclusion

Mistral Large 4 represents an important development in the open-weight AI landscape. Its trillion-parameter architecture, improved coding performance and cybersecurity capabilities make it a notable release for developers and businesses. While safety concerns and competition with closed models remain, Mistral’s latest launch strengthens Europe’s efforts to develop advanced AI technology and provide more accessible alternatives to proprietary systems.