Mistral releases one-trillion-parameter model to rival US and Chinese AI labs

Mistral AI released Mistral Large 4 on Tuesday, a one-trillion-parameter multimodal model the French lab is positioning as an alternative to both American closed models and Chinese open models. The company plans to release the weights publicly in roughly three weeks, once safety testing is complete.

The model, nicknamed Le Chonk in a nod to its trillion parameters, is not yet an open-weight release. For now, users can only access it through a public guardrail endpoint. That two-stage rollout reflects a bet Mistral is placing in an increasingly divided AI market: closed models that developers cannot inspect or relocate, and open-weight models that are, as Mistral sees it, often made in China.

The release follows what French president Macron has described as "a third way in AI." Europe's most prominent AI lab is leaning into that framing, presenting ML4 as a European frontier effort that can compete with the largest players on both sides of the open-closed divide.

Why the staged release matters

Mistral's core audience — enterprises and institutions — has grown more security-conscious in recent months, according to the company. That concern shapes how ML4 is being rolled out. Before the weights go public, Mistral intends to control access and vet usage through its endpoint and partner channels.

"In the meantime, we'll work with trusted partners and governments to make sure that the open-source weights can be used to defend, but not to [perform] malicious attacks," Mistral VP Science Pierre Stock told TechCrunch.

Stock also acknowledged the countervailing argument: an open-weight model is easier to audit. Once the weights ship, outside parties will be able to inspect the model directly rather than probe it through an API — a property that matters to institutional buyers who need to verify what a system actually does.

Trained on 4,000 GPUs

ML4 was trained entirely on Mistral's own compute. Stock said the effort used only 4,000 NVIDIA GPUs — "which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors."

That efficiency claim is central to Mistral's positioning. If a European lab can reach the frontier with a fraction of the training hardware used by larger rivals, the argument goes, the gap between open and closed ecosystems — and between Chinese, American, and European camps — becomes harder to sustain on compute grounds alone.

The timing also carries commercial weight. Samsung led Mistral's Series D last month at a €21 billion valuation (about $24.39 billion), and ASML, the Dutch semiconductor equipment giant, led its Series C. Both backers have direct interests in the domains Mistral says ML4 is optimized for.

What are ML4's target use cases?

According to Stock, ML4's optimized use cases include:

• Cybersecurity — a domain where Mistral's institutional customers have expressed mounting security concerns.
• Finance — a sector where multimodal capabilities can add value for Mistral's enterprise base.
• Chip design — a field core to two of Mistral's main backers: ASML and Samsung.

Stock said that thanks to focused training, the model could outperform closed models in specific areas that are key to Mistral's customers, not just match them on general benchmarks.

Can it beat the open-weight field?

Benchmark results for ML4 are still pending. Stock said Mistral hopes the model will be best in class among open-weight models, especially outside of China — though he did not limit that ambition to non-Chinese competitors.

The claim is untested until independent evaluations appear. But the stakes are clear: the open-weight leaderboard has been dominated by Chinese labs, and a European model competing at that tier would strengthen the case for the "third way" Macron described — an ecosystem that is neither closed and unplug-able nor exclusively Chinese-sourced.

Frontier lab or inference provider?

The release also serves a reputational purpose. When Mistral recently decided to host Chinese models, the company tried to convey that the move was not a pivot into becoming a mere inference provider. With Le Chonk in its corner, Mistral believes it should still be considered a frontier lab as well.

That distinction matters for Mistral's valuation story. A €21 billion valuation rests on the premise that the company builds frontier models, not merely serves them. ML4 — trained in-house, on Mistral's own 4,000 GPUs — is the evidence the company is offering.

The next test arrives in roughly three weeks, when the weights are scheduled to go public and outside evaluators can finally measure Le Chonk against both the closed American frontier and the open Chinese one.