Mistral has unveiled its latest AI model, nicknamed le Chonk, saying it’s the best open-weight model on the market for security – outside China, at least.

France-based Mistral positions itself as an alternative frontier AI developer to American companies, such as OpenAI and Anthropic, making it a viable option for sovereign AI.

The company unveiled a public preview of its next model, Mistral Large 4 (ML4), or as they also call it, le Chonk. The preview API is currently available, with full weights released later this month after security tests are complete.

“The model we’re actually announcing today is actually above the Chinese models on certain aspects, including cyber,” CEO Arthur Mensch told journalists per Reuters. “So the narrative that Europe cannot compete is something that is not true.”

What is le Chonk?

ML4 features one trillion parameters with 49 billion active parameters, according to Mistral.

“It is a significant milestone in our long-term investment across infrastructure, research, and product development: state-of-the-art performance in critical verticals, delivered through open weights, designed to give customers control over their AI,” the company said in a blog post, noting it can be run on private cloud or on-premise.

ML4 is multimodal, so it can work across documents, charts, and images, with Mistral saying it topped GPT-6-Astra on one visual grounding benchmark. It brings together instruction, reasoning, and agentic abilities into one model.

Mistral said it trained ML4 in partnership with key industries including engineering, finance, science, and the public sector. The model will eventually serve as the foundation for specialized Mistral models built for specific tasks and sectors.

Mistral also claimed leaps forward in math and scientific reasoning, making the new model a “strong research assistant… from the first question to the final result.”

Further improvements are expected via reinforcement learning post-training, helped by EU funding for European data centres.

“The reinforcement learning run behind this preview is still in flight, and the model is showing no signs of saturation — there is substantial headroom ahead,” Mistral added. “As we scale up training on our expanded infrastructure, we expect large and rapid improvements in the weeks and months to come.”

Controls for Security

Mistral claimed ML4 is among the leading global models in terms of cybersecurity capabilities. On the Artificial Analysis Cyber Index, an independent evaluation of how well AI models fix security flaws, ML4 secured a top-five place. It also leads in the open-weight model domain, albeit among those developed outside China.

Control is key to ML4’s success on this front, with Mistral noting that the open weights and self-deployment aspects of its models mean organizations can run security using their own policies, rather than being limited by their AI provider.

Mistral said it solved 93% of the exercises in Cybench, whereas closed models including Claude Opus 5.5 and GPT-6 Astra scored near zero because they refused to perform aspects of the task.

“Yet defending software often starts with proving that a flaw is real, exactly the kind of work safety filters in closed models can block,” Mistral added in the blog post. “This matters even more as threat actors increasingly jailbreak those same models to support offensive cyber activity: defenders need systems that can match those capabilities without being constrained by the same refusals.”

That said, Mistral stressed high scores in safety and security, pointing to the model’s “propensity to refuse malicious requests regarding cybersecurity.”

Sovereign AI

Mistral has made much of the fact that the company and its infrastructure are based in Europe, noting that ML4 was “trained from scratch” on 3,800 Grace Blackwell GPUs in Mistral’s own European data centres, with the public preview served on the same systems.

The model will still be available more widely around the world, but Mistral is also offering an entirely European deployment that it operates end-to-end, with no other digital service providers.

The company added: “Fun fact: a significant share of ML4’s training data was multilingual, spanning more than 160 languages, including every official language of the European Union.” With full weights set for imminent release and reinforcement learning training still ongoing, Mistral expects ML4’s capabilities to continue improving significantly in the months ahead.