Mistral ML4 Packs 1 Trillion Parameters, Open Weights Due in Three Weeks
Mistral unveils 1 trillion parameter ML4, trained on 4,000 Nvidia GPUs, with guarded access now and open weights planned after three weeks of testing.
Summary
French AI lab Mistral AI released Mistral Large 4 on Tuesday, October 6, 2026, presenting the multimodal model as President Macron’s “third way” between API controlled closed systems and open models often made in China. Nicknamed Le Chonk for its 1 trillion parameters, ML4 currently runs only through a public guardrail endpoint. Mistral plans to release its weights in three weeks after safety testing. Science Vice President Pierre Stock said trusted partners and governments will help enable defensive uses while limiting malicious attacks. Open weights would also improve auditing for security-conscious enterprises and institutions.
ML4 was trained entirely on Mistral’s compute using 4,000 Nvidia GPUs, two to three times fewer than Chinese competitors and significantly fewer than closed rivals. Benchmarks remain pending, but Mistral expects open-weight leadership, particularly outside China, and potential advantages over closed models in focused multimodal work spanning cybersecurity, finance and chip design. Chip design is central to two major backers: Dutch company ASML, which led Mistral’s Series C, and Samsung, which led its Series D the previous month at a €21 billion valuation, about $24.39 billion. After hosting Chinese models raised questions about an inference-provider pivot, Mistral argues ML4 reinforces its status as a frontier lab.
Positives
- ML4’s 1 trillion parameters give Mistral a large multimodal contender against American, Chinese, open and closed rivals.
- 4,000 Nvidia GPUs trained ML4, two to three times fewer than Chinese competitors and significantly fewer than closed-model developers.
- Open weights are scheduled within three weeks, potentially giving enterprises and institutions greater control and auditability.
- Cybersecurity, finance and chip design receive focused optimization aligned with Mistral customers and backers ASML and Samsung.
- Samsung’s Series D valued Mistral at €21 billion, about $24.39 billion, one month before ML4’s release.
Risks & concerns
- Benchmark results remain pending, leaving Mistral’s expected open-weight leadership and specialized advantages unverified.
- ML4 is not open weight at launch and remains accessible only through a public guardrail endpoint.
- Releasing 1 trillion parameter weights could enable malicious attacks despite planned safety testing and cooperation with governments.
- Security concerns are mounting among Mistral’s core enterprise and institutional customers.
- Hosting Chinese models created doubts over whether Mistral was shifting from frontier research toward inference services.