French AI company Mistral has released the model weights for its latest flagship system, Mistral Large 4 (ML4), internally known as “Le Chonk”. The company says the model delivers significant coding improvements and could strengthen cyber defences as Europe seeks greater control over its technology infrastructure.

Mistral describes ML4 as “the best open agentic coding model in the world”, giving businesses and governments greater control over AI tools for detecting, analysing and responding to threats across software and IT infrastructure.

Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram

The release follows a preview announced on 6 October. Since then, Mistral has continued reinforcement learning and tested the model with customers and partners in real-world environments.

“ML4 OSS showcases our science team’s ability to innovate at speed,” said Pierre Stock, Mistral’s vice president of science. The open-weight version builds on the preview with further training and performance gains, he added. Customer feedback was incorporated in less than three weeks, bringing two reasoning modes and improvements to efficiency and customisation.

Coding Performance And Cybersecurity Risks

Mistral says ML4 is competitive with leading open coding models and uses an average of 47% fewer tokens per task than China’s GLM-5.3. On Terminal-Bench 4.0, a benchmark for terminal-based coding tasks, its score rose from 28% to 41% after additional training.

The launch comes amid growing cyber risks in Europe, including a major PHP vulnerability reportedly identified with AI assistance and cyberattacks affecting Belgium’s finance and tax agency and Norway’s public sector.

AI can help security teams analyse code and infrastructure at scale, but it can also introduce new risks. Secure deployment, access controls and human oversight remain essential.

Why Open Weights Matter

Making model weights available allows organisations to run AI on their own infrastructure rather than rely solely on a provider’s hosted service. This can give businesses handling proprietary code or sensitive data greater control over information processing and system integration.

However, self-hosting requires sufficient computing capacity and technical expertise. Mistral also argues that on-premises deployment reduces the risk of third-party providers restricting or withdrawing access.

The issue has gained prominence as Europe seeks to reduce its 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. Restrictions were later 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’s AI Ambitions Face Infrastructure Constraints

Europe continues to lag behind the US and China in AI investment and computing resources. Mistral is among the region’s leading AI companies, but it competes in a market where US rivals have access to substantially more capital.

The company says ML4 was trained on 3,800 Nvidia Grace Blackwell GPUs at its European data centres. It argues that the model delivers state-of-the-art performance with a fraction of the computing resources and capital invested by closed-model developers such as OpenAI and Anthropic.

European data centres can give customers greater control over where workloads run, but reliance on Nvidia chips highlights continued dependence on foreign suppliers. Data residency and model ownership can improve operational autonomy, though neither guarantees complete technological independence.

“Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors,” said Arthur Mensch, Mistral’s CEO and co-founder.

Mensch said Mistral had raised roughly $6 billion in three years, compared with the hundreds of billions raised by other AI labs developing similar capabilities. The company aims to narrow the gap with better-funded competitors by improving computing efficiency and making its research widely available.

A Trillion-Parameter Model With Selective Computing

ML4 contains 1.05 trillion parameters but activates approximately 52 billion for each task, according to Mistral. This architecture allows it to draw on a broad range of capabilities without activating every parameter for every request, potentially improving efficiency.

Users can select low, medium or high reasoning settings depending on the complexity of a task.

Mistral says ML4 can convert two-dimensional blueprints into three-dimensional models using computer-aided design software, analyse large datasets for financial decisions, and interpret satellite and aerial imagery to identify wildfires, assess crop health and monitor risks to power lines.

What Comes Next

Mistral says ML4 is part of a new generation of multimodal models, with further releases planned in the coming months. It is available through the company’s API, and its weights can be downloaded from Hugging Face.

For European organisations, ML4 offers a way to gain greater control over AI deployment, but its reliance on advanced foreign-made chips illustrates the limits of that autonomy. Achieving lasting technological independence will require not only accessible models, but also secure infrastructure, skilled technical teams and a stronger technology supply chain.