Mistral Previews Large 4, a Trillion-Parameter Open-Weight Model

French AI lab Mistral AI has opened a public preview of Mistral Large 4, a one-trillion-parameter multimodal model it calls the strongest open-weight system built outside China. The company unveiled the model on Tuesday, nicknamed "Le Chonk," and said its weights will ship by the end of October, after red-teaming with cybersecurity partners and state authorities.

What Mistral announced

ML4 carries 49 billion active parameters and was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs inside Mistral's own European data centers. The preview is live today through Mistral Studio and the company's API; the downloadable weights and final license are still pending.

Mistral is aiming the model at enterprise workloads rather than general chat. It reports state-of-the-art open-model results in cybersecurity, finance, and law, and says ML4 beats even frontier closed models on tasks such as visual grounding. On the Artificial Analysis Cyber Index it ranks in the global top five and leads open-weight models developed outside China by a wide margin, scoring 82% on a test that asks a model to reproduce and then patch a real vulnerability. The company says it used two to three times fewer GPUs than its Chinese rivals.

Why the open-weight timing matters

The three-week gap between preview and weights is deliberate. Mistral VP of Science Pierre Stock told TechCrunch the delay lets the company work with trusted partners and governments so the open weights "can be used to defend, but not to perform malicious attacks." For security-conscious enterprises, that tradeoff is the whole pitch: an open model can be audited and self-hosted, but it is also harder to unplug once deployed.

The move puts Mistral between two camps. Closed American models offer polish but can be revoked; open Chinese models are cheap and capable but raise sovereignty and compliance questions for European buyers. Mistral is betting that a European-trained, multilingual model, with a significant share of its training data spanning more than 160 languages including every official EU language, wins the customers stuck in the middle.

Builder impact

For teams shipping AI products, ML4 signals that frontier-grade open weights are no longer a uniquely Chinese phenomenon. The preview API lets developers benchmark the model against their current stack now, before committing to a self-hosted deployment when the weights land. Watch the cybersecurity angle in particular: because provider-level refusals can block legitimate vulnerability research, an open model with strong cyber scores and self-hosting is attractive to security teams that currently work around closed-model guardrails.

Mistral says ML4 will also become the foundation for a new generation of specialized models, suggesting more vertical releases are coming.

Sources

• Mistral: Introducing Mistral Large 4
• TechCrunch: Mistral's new 1T model aims to leapfrog closed and open rivals