Mistral’s Le Chonk Enters the Open AI Race

Mistral has introduced Mistral Large 4 (ML4), also called “le Chonk,” a 1-trillion-parameter AI model that the French company says is among the world’s strongest open models and the most capable developed outside China by a substantial margin. Mistral says the model performs particularly well in cybersecurity, coding, manufacturing, finance, and multimodal tasks. It is initially available in preview to developers, cybersecurity leaders, and government authorities, with a broader release planned for later in the month. The launch comes as open AI models have become a focal point in competition between the United States and China. Unlike closed systems from companies such as OpenAI and Anthropic, open models can be modified and hosted by their users. Chinese models have led the field and are gaining adoption internationally, prompting Western developers to work to close the capability gap. Mistral says ML4 was trained for two months on 4,000 Nvidia Grace Blackwell GPUs in its European data centers. The company expects the model to rank among the leading open-weight systems on combined benchmark results once its core parameters are released. However, the article notes that ML4 remains behind frontier models in some areas, including coding. Co-founder and chief scientist Guillaume Lample said the company expects capabilities to improve as it expands training capacity following its September Series D funding round. Mistral raised 3 billion euros, or $3.4 billion, at a 21 billion euro valuation, reinforcing its position as one of Europe’s best-funded AI startups. The article also cites a cybersecurity incident involving Hugging Face, which used Chinese company Z.ai’s GLM 5.2 after U.S. closed models’ guardrails blocked its defensive activity. Separately, Nvidia-backed Reflection AI announced an open model it said matched GLM 5.2, which had since been surpassed by GLM 5.3.

French AI company Mistral has announced Le Chonk, officially named Mistral Large 4, its largest open-weight AI model to date. The model is aimed at users who want to keep their data private: unlike closed models from companies such as Anthropic and OpenAI, its downloadable weights let users run and customize the software on their own servers and hardware. Mistral says the model will be publicly available on October 27. Cybersecurity experts and government authorities will also receive access to a version with fewer safety restrictions so they can test its capabilities. The launch is significant for Mistral, which is three years old and is regarded as Europe’s leading AI company. It is the company’s first new model release since May. Mistral says Le Chonk outperforms many competing open-weight models. CEO Arthur Mensch went further at an event in Abu Dhabi, saying the model is “actually above the Chinese models” in some areas, including cyber capabilities. He did not identify the Chinese models he was comparing it with. The announcement comes after Mistral raised €3 billion ($3.4 billion) last month, described in the article as Europe’s largest technology fundraising. The company has presented itself as a European alternative to major US AI firms, particularly as those companies have recently warned about the dangers of AI. The article places the launch within Europe’s broader effort to build AI capabilities and achieve greater digital independence, while noting that open-weight models are common among many Chinese competitors and less characteristic of leading US models.

French AI company Mistral AI has launched a public preview of ML4, a frontier model it says will be the most capable open-weight AI model developed outside China once its weights are released. Nicknamed “Le Chonk,” ML4 is designed with a particular focus on cybersecurity and defence. Mistral says these capabilities could help businesses and governments defend against cyber threats, including attacks enabled by people jailbreaking closed models. The model’s launch comes amid growing competition between Chinese open-weight developers and American companies that often keep their most capable systems closed, as well as concern about AI agents acting autonomously and hacking websites. Mistral co-founder and chief scientist Guillaume Lample described ML4 as a new generation of models. Vice president of science Pierre Stock said the company is closing the gap with competitors and that ML4 is stronger than some Chinese models released earlier in the year. Mistral also claims strong performance in manufacturing, finance, multimodal use, coding, grounding and semiconductor benchmarks. The company cautions that training is not complete: reinforcement learning is expected to finish in the coming days, and the model should continue to improve. Its performance claims have not been independently verified in the article. Mistral says ML4 was trained from scratch over two months on 4,000 NVIDIA Grace Blackwell GPUs in its European data centres. It supports more than 160 languages, including all official EU languages, and both Latin and non-Latin scripts. The model is available through Mistral’s API on Mistral Studio, while the company plans to release its weights on October 27. Developers, cybersecurity leaders and public authorities can test it before then. The launch also follows Mistral’s announcement of a €3 billion funding round at a valuation above €21 billion, which the company calls the largest equity round by a European technology company. Overall, Mistral is positioning ML4 as a European, sovereignty-focused alternative in an increasingly geopolitical open-weight AI race.