Mistral launches Le Chonk, a trillion-parameter open-weight model targeting code and cybersecurity

French AI lab Mistral unveiled Mistral Large 4 yesterday, a trillion-parameter model known internally as Le Chonk. The model is available now in preview; a final release is expected by the end of the month. The company describes it as the most advanced open-weight model developed outside China and says its performance is "very close" to that of leading proprietary systems. Unlike Chinese laboratories accused by the U.S. government of using distillation — training smaller models on the outputs of larger ones — to close the gap with OpenAI and Anthropic, Mistral says the model was trained from scratch.

The model is not aimed solely at general-purpose use. It has been optimized for code, cyber defense, and specialized tasks in manufacturing, finance, electrical engineering and other niche domains. Co-founder and chief scientist Guillaume Lample said the major labs focus less on those areas, and that is where the room for improvement lies. From a business standpoint, open-weight models are significantly cheaper to operate — the cost amounts to compute alone — and Mistral is betting that narrowing the performance gap will remove the final barriers to enterprise adoption of open source.

Mistral's business model differs from its U.S. rivals. Instead of charging a premium for access to closed models, the company bills per use for running models on its own cloud and sells engineering services to customize models for customers. That strategy is starting to show results: in September the company raised $3.3 billion at a $24 billion valuation, the largest funding round ever for a European tech company. Revenue has reportedly jumped 20-fold over the past year.

Mistral's rise comes against a backdrop of growing friction between the United States and its European allies — from tariff policy and Greenland to regulation of American tech giants. In June the U.S. administration imposed temporary restrictions on distribution of OpenAI and Anthropic models over concerns they could be used in sophisticated cyberattacks. Since then, cases have emerged in which American models "broke free" of their guardrails and targeted foreign companies and government institutions, fueling a prolonged debate over model-release regulation. The White House has reportedly asked U.S. labs not to share unpublished models even with the UK AI Safety Institute, which previously assisted in risk assessment.

The reminder that a U.S. administration can unilaterally cut off access to frontier models has created an opening for a European lab with open-weight offerings. Andrea Renda, research director at the Centre for European Policy Studies, said in July that the EU's tech-sovereignty strategy combined with rising U.S. hostility had produced a "magic formula" that puts Mistral in a comfortable position, even though its performance had not been especially impressive until now. For its part, Mistral says it does not want to be seen as serving only its home market.