Mistral Large 4: 1T-Parameter "Le Chonk" Targets Rivals

French AI lab Mistral AI released Mistral Large 4 (ML4) on Tuesday. It is a large multimodal model, meaning it handles more than one type of input, and Mistral wants it to compete with both American and Chinese labs. The release also shows that Europe still intends to compete at the top end of AI development.

Mistral frames the launch around an idea that French president Emmanuel Macron has called "a third way in AI." The market is increasingly divided between closed models, which their owners can switch off, and open models, many of which come from China. Mistral is pitching ML4 as an alternative to both.

A trillion parameters, weights still to come

ML4 has 1 trillion parameters, which explains its internal nickname, Le Chonk. It is not yet an open-weight model, though. "Open weight" means the trained model parameters are published so others can download, run and inspect them.

For now, the only way to use ML4 is through a public endpoint with guardrails in place. Mistral says it will release the weights in three weeks, once safety testing is finished.

Pierre Stock, Mistral's VP Science, told TechCrunch the company will spend that time working "with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks."

The staged rollout reflects growing security worries. According to Stock, those concerns have increased in recent months, especially among enterprises and institutions, which are Mistral's core customers. Stock also argued that open weights have a security benefit of their own, because an open model is easier to audit.

Trained on a smaller GPU budget

Mistral trained ML4 entirely on its own compute, using 4,000 Nvidia GPUs. Stock said that is "two to three times less than our Chinese competitors, and significantly less than the closed source competitors."

That number matters. It suggests Mistral is presenting efficiency, rather than raw scale, as its edge against rivals with far more hardware.

Where Mistral expects ML4 to win

Benchmark results have not been published yet, so all performance claims are still unverified. Stock said Mistral hopes ML4 will be the best open-weight model available, especially outside China, though he did not limit the ambition to that market.

The company is also aiming at specific domains. Because the training was focused, Mistral thinks ML4 could beat closed models in areas that matter most to its customers, particularly where multimodal features are useful. Stock listed three target use cases:

• Cybersecurity
• Finance
• Chip design

Chip design connects directly to Mistral's investors. Dutch chip equipment maker ASML led Mistral's Series C round. Samsung led its Series D last month, which valued the company at €21 billion (about $24.39 billion).

More than an inference provider

There is also a positioning issue. Around its latest funding round, Mistral said its decision to host Chinese models did not mean it was becoming a simple inference provider, a company that only runs other developers' models. With Le Chonk, Mistral is arguing that it should still be counted as a frontier lab that trains its own leading models.

The Bigger Picture

For readers following the open-model race, ML4 is a direct challenge to the idea that serious open-weight models now come mainly from China. US startups are making similar moves, as seen with Reflection's Beam model. In Europe, Aleph Alpha has also pushed open-weight releases aimed at enterprise users, which suggests an alternative to US and Chinese labs is starting to take shape in the region.

The three-week gate before releasing the weights may turn out to be the more significant part of the launch. Recent results show open models getting stronger at offensive security tasks. Mistral's plan to test with governments before publishing could become a template for how large open models are released.

Three things are worth watching. First, whether independent benchmarks back up Mistral's best-in-class claims. Second, whether the weights actually ship on schedule. Third, whether ML4 proves useful in chip design for partners such as ASML and Samsung.