Many of the most powerful artificial intelligence (AI) models remain under the control of the companies that developed them, limiting how freely other businesses can run or customise the technology.
But French AI startup Mistral is looking to challenge that arrangement with Mistral Large 4, a powerful new model it plans to make available for others to operate independently.
Announced on October 6 and nicknamed “Le Chonk,” the model has roughly one trillion parameters, making it Mistral’s largest and most capable model yet.
It is currently available through a public preview of the company’s API, with its downloadable model weights scheduled for release on October 27.
What Makes Mistral Large 4 Different
Parameters are the numerical values an AI model learns during training. Although Large 4 contains more than a trillion, it uses only about 50 billion at a time through a system called mixture of experts, which selects different parts of the model for different tasks.
This allows Mistral to build a large model without using all its parameters for every response.
According to the French company, Large 4 can also analyse images and documents, work with large amounts of text, and handle more than 160 languages.
The company says the model performs strongly in software coding, cybersecurity, financial analysis, and engineering. It reported a 61.7% score on the DeepSWE coding benchmark and 82% on a test involving the reproduction and repair of real software vulnerabilities.
However, these figures come from Mistral’s published evaluations and do not establish that it outperforms every competing model.
Why Open Weights Matter
For developers, Large 4’s weights is the hook, as it offers the trained information that determines how the model works.
Once available, organisations with sufficient computing resources could deploy it on their own infrastructure and adapt it to specific business needs rather than relying entirely on an external AI service.
This could be valuable to banks, manufacturers, and cybersecurity teams handling sensitive data or requiring greater control over their systems.
Mistral says it trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres, which reflects its wider push for AI infrastructure that organisations can control.
Mistral co-founder Guillaume Lample also highlighted the importance of this independence, saying, “If you use a closed model, there is no guarantee it will still be there tomorrow.”
What This Means for the Open-Source AI Market
Large 4 is entering a market where Chinese developers, including DeepSeek, Alibaba, and Z.ai, have become major competitors in downloadable AI models. Mistral argues its release can give customers another powerful option from Europe.
However, open-weight does not automatically mean fully open-source, because making a model’s weights available does not necessarily include releasing its training data or development code.
Mistral has also not published the final licensing terms for Large 4, meaning the extent to which developers can freely modify and distribute it remains uncertain. Plus, running a model of this size will require substantial computing resources.
But ahead of the planned release, Mistral is testing Large 4 with cybersecurity specialists and government authorities to assess potential safety risks.
For now, Large 4 aims to strengthen competition in the AI industry by bringing another advanced model closer to public availability. And its wider impact will depend on the final licensing terms, independent performance tests, and how easily developers and businesses can put the model to work
