Mistral AI has released Mistral Large 4, a new multimodal model built around one trillion parameters and positioned as a European alternative to both closed American systems and open models developed in China. The French company introduced the model on Tuesday, according to TechCrunch, framing the launch as evidence that Europe remains part of the frontier AI race.

The model, also nicknamed “Le Chonk,” is not yet available as an open-weight release. For now, TechCrunch reported, users can access it through a public endpoint protected by guardrails. Mistral plans to publish the model’s weights in three weeks, after completing safety testing. That timeline makes the current launch an initial access phase rather than the final form of the release.

Guarded model access shown before a future open-weight release.
Mistral Large 4 is currently available through a protected endpoint, with model weights planned after safety testing.

Mistral vice president of science Pierre Stock told TechCrunch that the company will use the intervening period to work with trusted partners and governments. The stated goal is to make the eventual weights useful for defensive applications while reducing their use in malicious attacks. Stock also argued that open-weight models can be easier to audit, an important consideration for the enterprises and institutions Mistral serves.

The company says Mistral Large 4 was trained entirely on its own computing infrastructure using 4,000 Nvidia GPUs. Stock claimed that this was two to three times fewer GPUs than Chinese competitors used and significantly fewer than closed-model rivals. TechCrunch noted that benchmark results are still pending, so those efficiency and performance comparisons remain company assertions rather than independently demonstrated conclusions.

Abstract GPU cluster powering a multimodal AI model.
Mistral says it trained the model on its own infrastructure using 4,000 Nvidia GPUs and optimized it for specialized enterprise work.

Mistral hopes the model will rank among the leading open-weight systems, particularly outside China, once its weights are released. Stock said focused training could also let it outperform closed models in specific customer-facing areas. The company identified cybersecurity and finance as optimized use cases, along with chip design, where multimodal capabilities may be valuable.

Chip design also connects the model to two of Mistral’s major financial backers. TechCrunch reported that Dutch semiconductor-equipment company ASML led Mistral’s Series C, while Samsung led its Series D in September at a €21 billion valuation, approximately $24.39 billion. Mistral’s emphasis on that workload therefore aligns the model with strategically important customers and investors as well as the broader competitive narrative.

The release is also intended to reinforce Mistral’s identity as a frontier model developer. The company had previously faced questions about whether hosting outside models signaled a shift toward becoming mainly an inference provider. With Mistral Large 4, it is making the opposite case: that it can develop a very large multimodal system on European infrastructure, release it behind safeguards first, and move toward open weights after testing. Whether the model fulfills its competitive ambitions will depend on the forthcoming benchmarks and the promised weight release.