Mistral releases latest AI model for download

The French AI company says its Large 4 model offers advanced coding and cyber-defence capabilities and can be run on an organisation’s own computers and servers.

Mistral has released the weights for its latest large flagship AI model, Mistral Large 4 (ML4), nicknamed “Le Chonk”, claiming improvements in coding and cyber defence as Europe seeks to build alternatives to US and Chinese technology.

The French firm claims the model is “the best open agentic coding model in the world”, offering enterprises and governments “frontier AI cyber defence capabilities they can fully own and control” including the ability to “detect, analyse and respond to cyberthreats across code and IT infrastructure“.

The release follows a preview of ML4 on 6 October. Since then, Mistral has spent the last three weeks reinforcing the model’s learning, testing it with customers and partners under real-world conditions.

“ML4 OSS showcases our science team’s ability to innovate at speed. Building on the ML4 Preview, the open weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation,” said Pierre Stock, Vice President of Science at Mistral.

“In under three weeks within the same reinforcement learning run, we integrated customer feedback by introducing two additional reasoning modes, while making ML4 more efficient and customisable,” he added.

Mistral says the latest release of ML4 rivals other leading open models for coding while using 47% fewer tokens per task on average than Chinese model GLM-5.3. Its score on Terminal-Bench 4.0, which tests AI agents’ ability to perform coding tasks in a terminal, also rose from 28% to 41% following further training.

The release also comes as Europe faces growing cybersecurity threats, such as the recent PHP programming security flaw, discovered using AI, as well as the recent cyber-attacks on the Belgian and Norwegian governments.

Making the model’s weights available allows organisations to download and run it on their own infrastructure, rather than relying exclusively on an external provider’s hosted service.

This could be particularly useful for companies handling sensitive information, such as proprietary software code, that they do not want to share with an external AI service.

Organisations would have greater control over where the model runs and how it is integrated into their systems, although they would still need suitable hardware and expertise to operate it securely.

Mistral says organisations running ML4 on their own infrastructure would also not face the risk of a third-party provider restricting access to or disabling the model.

The move comes amid concerns about European dependence on US AI providers. In June, the Trump administration ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models for foreign nationals.

The restrictions were subsequently lifted for Fable 5, while access to Mythos 5 remained limited to selected organisations. The episode highlighted how decisions made outside Europe can affect access to advanced AI services.

Europe continues to lag behind the US and China in the development of large AI models and the resources needed to build them.

Mistral is one of Europe’s leading AI developers, but its reported valuation of around $24 billion (€20.9 billion) is far below those of US rivals OpenAI and Anthropic. [Insert verified, dated valuations and

Mistral argues that despite the disparity in resources, “ML4 is now the best open model built in the West”.

The model was trained on 3,800 of US company Nvidia’s Grace Blackwell GPUs at Mistral’s European data centres. The company says ML4 achieves “state-of-the-art performance” while mobilising “a fraction of the compute and capital committed by closed labs” such as OpenAI and Anthropic.

The model’s European infrastructure could give customers greater operational control, but its reliance on Nvidia chips illustrates Europe’s continued dependence on foreign suppliers for key parts of the AI supply chain.

“Openoncentrated in the hands of a few actors,” said Arthur Mensch, CEO and co-founder of Mistral

“Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion.”

“We have taken the lead globally in compute efficiency and have widened the gap between Mistral and other labs in that area. Our approach is to deliver quality science and put the results in everyone’s hands,” he added.

ML4 has 1.05 trillion parameters in total, but activates around 52 billion This means it does not use all its parameters for every task, helping it process requests more efficiently

The model also offers three reasoning modes — low, medium and high — allowing users to adjust the depth of reasoning to suit the task, from everyday requests to more complex problems.

Mistral says ML4 can turn 2D blueprints into usable 3D models using computer-aided design (CAD) software, analyse large volumes of data to support financial decision-making and use satellite and aerial imagery to detect wildfires, track crop health and monitor risks to power lines.

The French firm says the release of ML4 is part of a new generation of models spanning all modalities and that they plan to release more models from this generation in the coming months.

The model is now available through Mistral’s API, while its weights can be downloaded from Hugging Face.

Source: www.euronews.com