"Le Chonk": Mistral's new trillion-parameter model is the strongest open model outside China — but only eighth overall
On October 6, 2026, the Paris-based company Mistral AI launched a public preview of Mistral Large 4, a multimodal model with roughly one trillion parameters and the code name "Le Chonk". The company presents it as a return to the global frontier for open weights, and it is the strongest open model outside China on the independent Intelligence Index from Artificial Analysis. But the same ranking shows the model is only eighth among open-weight models — behind seven Chinese systems — and more than 19 points behind the best closed models. The weights are to be released on October 27 under a custom Mistral license, whose terms remain unknown.
What was launched, and how the model is built
Mistral Large 4, abbreviated ML4, is a sparse mixture-of-experts model with one trillion total parameters, of which 49 billion are active per token, according to VentureBeat. The focus areas, according to the company, are coding, cybersecurity, finance, manufacturing and visual grounding. The documentation states a context window of one million tokens, which makes it possible to process large amounts of data in a single request (Anadolu Agency, b798d9c1).
According to VentureBeat, the model was trained from scratch over roughly two months on Nvidia Grace Blackwell GPUs in Mistral's own European data centres. Vice President for Science Pierre Stock told TechCrunch that this involved 4,000 GPUs, "which is two to three times fewer than our Chinese competitors" (TechCrunch, cbc0653c). According to Stock, this is significantly fewer than what closed-source competitors use, but how many fewer is not specified. Note that some outlets, including WOWTALE and Anadolu Agency, report the figure as 3,800 — the difference is unresolved in available reporting.
Chief Scientist and co-founder Guillaume Lample called ML4 "a new generation of models" at a press conference covered by Euronews (Euronews via MSN, 5023e66b). The code name "Le Chonk" means "the big, chunky one", according to the same source.
The plan is to make the model available through Mistral's API and publish the weights on October 27, after a testing period of roughly three weeks with developers, cybersecurity leaders and government authorities. The weights are expected under a custom Mistral license, but the exact license terms have not been announced (VentureBeat, 15b0892d).
What the independent test actually shows
The first independent evaluation confirms half of Mistral's presentation. On Artificial Analysis' Intelligence Index — one of the most widely followed benchmark services for AI models — Mistral Large 4 scores 38.4 points. That is a massive leap from its predecessors: Mistral Medium 3.5 scored 14 points and Large 3 scored 9 (Trending Topics, 22f96a10).
But that is where the frontier story stops. At 38.4 points, Large 4 is eighth among open-weight models, behind seven Chinese systems, with Xiaomi's MiMo-V2.6-Pro at the top with 46.3 points. The gap to the leading closed models is larger: Anthropic Claude Opus 5.5 scores 57.6 points in its strongest configuration, OpenAI GPT-6 Astra 52.7 and Google Gemini 4 Argon 52.6. That places Large 4 more than 19 points behind the world's best — roughly two-thirds of Claude Opus 5.5's score, according to Trending Topics' review of the ranking.
One technical detail is worth noting: because the weights have not yet been released, Artificial Analysis is currently treating the preview as a proprietary model. The ranking among open models therefore only becomes final after October 27.
The cyber niche: strong on paper, but vendor-reported numbers
Mistral's most striking claims concern cybersecurity, and they are not yet independently verified. According to WOWTALE's coverage of the company's own figures, Large 4 scores among the top five models globally on the Artificial Analysis Cyber Index, which measures how well a model finds and fixes real software vulnerabilities. The model reportedly achieved 82 percent on a test asking the model to reproduce a real vulnerability in open-source software and then patch it — stated as the highest of any model. It is also reported to have solved 93 percent of Cybench, a set of 40 tasks drawn from security competitions (WOWTALE, a90e3534). Forkast News points out that these figures are vendor-supplied and not independently verified (Forkast via Yahoo Finance, 89aa15fd).
The bet has a deliberate strategic logic. Lample has argued that cyber-defence capabilities let companies and governments defend themselves against threat actors who break through closed models via jailbreaks. Stock elaborated to TechCrunch on what the company wants to achieve before the weight release: the weights should be usable "to defend, but not to carry out malicious attacks", he said, adding that Mistral will in the meantime work with trusted partners and authorities to ensure this (TechCrunch, cbc0653c). It is still unclear how this balancing is to be enforced in practice — it will depend largely on the license terms, which have not been announced.
Pricing and capital: conflicting reporting
On pricing, the sources disagree. Forkast News via Yahoo Finance reports that Mistral is pricing ML4 at $0.68 per million input tokens and $2.09 for output, undercutting the approximate $2/$10 floor established by OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon (89aa15fd). VentureBeat, by contrast, writes that Mistral did not state API pricing in the materials it reviewed (15b0892d). No explanation for the difference between the two accounts appears in the available source material, and the price should therefore be treated as unconfirmed until Mistral itself confirms it.
Whatever the pricing, Mistral is now operating from a new financial position. Large 4 is the first product release after the €3 billion Series D round led by Samsung Electronics in September, which according to WOWTALE values the company at over €21 billion — described as the largest equity round ever raised by a European technology company (WOWTALE, a90e3534).
What to watch ahead of October 27
Three factors will determine whether the launch delivers what it promises:
The license terms. Only a "custom Mistral license" is confirmed. Since the model is positioned around cyber defence, the terms for commercial and security-related use will be what actually distinguishes ML4 from a fully open system — and they remain unknown.
Verification of the cyber numbers. The reported 82 and 93 percent figures come from Mistral. Once the weights are public, third parties can test them directly; the same applies to the claim that closed top models scored near zero on the vulnerability test because they refuse.
Clarification of the conflicting figures. The GPU count (4,000 according to Stock himself, 3,800 at two other outlets) and the API price have both been reported differently. The efficiency story — frontier-level performance on a fraction of the compute budget — is central to Mistral's argument, and it deserves precise numbers.
The fair summary of today's picture is calibrated: Mistral has delivered a large efficiency leap with European infrastructure, and the independent ranking confirms it as a genuine leap. But it also places the model behind seven Chinese systems and well behind the frontier among closed models. "Back at the frontier" is, as of today, a claim about direction — not about position.