Mistral launches trillion-parameter Large 4, trained on just 4,000 GPUs

Mistral released Mistral Large 4 on Tuesday, a 1-trillion-parameter multimodal model the Paris-based AI lab says it trained using roughly 4,000 Nvidia GPUs — a sliver of the compute budgets that rival labs are believed to spend on frontier-scale training runs. The model, nicknamed "Le Chonk" internally, is live now through a public guardrail endpoint, with open-weight release planned in about three weeks pending safety testing.

The launch lands in the middle of a widening three-way split in the AI industry: closed, proprietary systems from American labs, open-weight models from Chinese developers such as DeepSeek and Alibaba's Qwen, and a shrinking pool of Western open-weight alternatives. Mistral has positioned itself squarely in that third lane since its founding in 2023, and Large 4 is its most direct attempt yet to argue that open models trained in Europe can compete with both camps rather than trail them.

Pierre Stock, Mistral's VP of science, said the training run used “two to three times less” compute than the company believes its Chinese competitors required, and “significantly less” than closed-source rivals. Mistral has tuned Large 4 specifically for cybersecurity, finance, and chip-design workloads — enterprise niches where the company is courting paying customers rather than chasing general chatbot benchmarks. The company says it will work with “trusted partners and governments” on responsible use of the open weights once they ship, a line aimed at easing concerns that an unrestricted trillion-parameter model could be repurposed for offensive cyber or weapons-design use.

Independent benchmark results are not yet public, so Mistral's claim to be “best in class among open-weight models” remains unverified outside the company's own testing. If the efficiency numbers hold up under scrutiny, though, Large 4 would strengthen the case that frontier-scale models no longer require the enormous GPU clusters that OpenAI, Anthropic, and Google have built their reputations on — a claim with real stakes for how much compute capacity the rest of the industry believes it needs to buy.

As first reported by TechCrunch, Mistral did not disclose pricing for the guardrail endpoint ahead of the open-weight release.