Mistral AI released Mistral Large 4 on Tuesday, a large multimodal model the French lab is positioning as an alternative to both closed American systems and open models that largely originate in China. The model is nicknamed Le Chonk in reference to its one trillion parameters, according to the report. It is not yet an open-weight release: for now it can only be reached through a public guardrail endpoint, with weights slated to follow in roughly three weeks once safety testing concludes.
Mistral VP Science Pierre Stock told TechCrunch that during the interim period the company will work with trusted partners and governments so that the open-source weights can be used to defend rather than to carry out malicious attacks. The evidence does not state which partners or governments are involved, nor what criteria would determine that the weights are safe to publish.
Mistral's stated ambition is to be best in class among open-weight models, particularly outside China, though Stock said the company is not limiting that goal to non-Chinese competitors. Benchmark results were still pending at the time of the report, so the claim is a target rather than a demonstrated result. Nothing in the supplied material shows measured performance against named rivals, and the comparison to closed and open competitors therefore rests on Mistral's own positioning.
On compute, Stock said ML4 was trained entirely on Mistral's own infrastructure using 4,000 Nvidia GPUs, which he described as two to three times fewer than Chinese competitors and significantly fewer than closed-source competitors. That is a company-supplied figure about training scale, not an independently verified efficiency measurement, and the evidence does not include training duration, data composition or total cost.
According to Stock, cybersecurity and finance rank among ML4's optimized use cases, alongside chip design. The report connects that chip design angle to two of Mistral's principal backers. One is Dutch giant ASML, which led its Series C. The other is Samsung, which last month led its Series D at a €21 billion valuation (about $24.39 billion). The evidence does not state that either backer helped train ML4 or put it into deployment.
The report notes that French president Macron described the effort as a third way in AI, language that places ML4 between closed models that can be unplugged and open models often made in China. That is context for how the company markets the model; it is not evidence about the model's capabilities.
The security rationale cuts both ways in Mistral's account. Stock argued that an open-weight model is easier to audit, while the report notes that security concerns have been mounting among Mistral's core audience of enterprises and institutions. The company's answer is a phased handoff: guardrailed access first, weights later, with partners and governments involved in the interval. Whether that process produces a materially different risk profile than a direct open release is not established in the evidence.
The report also says the company previously tried to convey that hosting Chinese models was not a pivot into becoming a mere inference provider, and that with Le Chonk in hand it believes it should still be counted as a frontier lab. The evidence does not detail what hosting arrangement that refers to or when it occurred.
For freelancers, designers and developers, the practical near-term picture is constrained. Until weights ship, the only documented access path is the public guardrail endpoint, which means teams cannot self-host, fine-tune or audit the model locally, and any workflow that depends on open weights has to wait or use something else. The three-week window is a stated plan, not a commitment with a published date, and it is contingent on safety testing finishing.
The multimodal capability is the other element worth tracking. Mistral describes ML4 as a large multimodal model and says multimodal capabilities can add value in the areas it has optimized for, but the evidence does not specify which modalities are supported, what context limits apply, or how the guardrail endpoint restricts inputs and outputs. Those details matter more than the parameter count for most integration decisions.
Cost and licensing are unresolved in the supplied material. There is no pricing for the endpoint, no license terms for the eventual weights, and no statement about whether the open release will carry usage restrictions tied to the safety work Mistral describes. Developers evaluating ML4 against alternatives should treat those as open questions rather than assume an open-source outcome.
The competitive claim is the hardest to assess. Mistral says focused training could let ML4 outperform closed models in specific areas that matter to its customers, but with benchmarks pending, that remains a hypothesis. The evidence contains no third-party evaluation, no head-to-head results and no independent confirmation of the GPU count or the efficiency comparison.
What remains unknown is substantial: benchmark numbers, endpoint pricing, license terms, supported modalities, context limits, and whether the three-week timeline holds. The evidence also does not say which governments or trusted partners are involved, or what criteria would determine that the weights are safe to publish. Until those gaps close, ML4 is best treated as an announced model with limited access rather than a settled option for production work.