Mistral is launching a public preview of Mistral Large 4, a one-trillion-parameter multimodal model code-named “Le Chonk” that the Paris AI company says pushes it back toward the global open-weight frontier — with a particular focus on coding, cybersecurity, finance, manufacturing and visual grounding.
The model, abbreviated ML4, has 49 billion active parameters and was trained from scratch over roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral’s own European data centers.
Mistral says it trained the model across more than 160 languages, including every official language of the European Union.
The company plans to make it available through Mistral’s API and to publish the model weights on Oct. 27 after a roughly three-week testing period with developers, cybersecurity leaders and government authorities.
The weights are expected under a custom Mistral license; the company did not provide API pricing in the materials reviewed by VentureBeat.
That staged rollout is central to how Mistral is positioning the release. ML4 is not just another general-purpose model: the company is pitching it as a foundation enterprises and governments can eventually run and customize themselves, including on sovereign infrastructure and with zero-data-retention options. Mistral says the preview period will also give it time to continue reinforcement learning and tune the final checkpoint before the weights go live.
“ML4 is at the frontier of open weight models,” Guillaume Lample, Mistral co-founder and chief scientist, said in materials shared with VentureBeat.
Lample said Mistral expects its capabilities to improve as reinforcement learning concludes and the company expands training capacity. Mistral says Lample has scaled its science team from three researchers to roughly 300.
From ‘Le Chaton Fat’ meme to a real trillion-parameter model
For anyone who follows the increasingly strange online culture surrounding frontier AI models, ML4 may sound like the punchline to a joke that started months ago.
In June, a fictional Mistral model called “Le Chaton Fat” went viral across X and Reddit, complete with fake benchmark charts and increasingly preposterous specifications for a supposedly enormous French model that would leapfrog its American and Chinese competitors. Business Insider reported that the joke spread in June, with some versions claiming the nonexistent model contained more than 30 trillion parameters and others advertising such capabilities as “1,000 meows per second.” Mistral CEO Arthur Mensch eventually joined in, writing on X, “It’s actually le gros chaton.”
Beneath the shitposting was a real expectation that Mistral was preparing something much larger. By July, TechCrunch reported on anticipation around a coming large open-weight Mistral model and noted that Mensch and Mistral investor Marc Andreessen had amplified the Le Chaton Fat jokes.
Now Mistral actually has a trillion-parameter model — and it appears the company decided not to let the joke go to waste. Internally, ML4 carries the nickname “Le Chonk,” another play on internet slang for an exceptionally large cat.
When VentureBeat asked Lample whether ML4 was effectively the model that the Le Chaton Fat meme had anticipated, he said Mistral had enjoyed the meme and suggested ML4 could be viewed as an initial version of the idea, with still larger models to come. Mistral executives said the Le Chonk name deliberately nods to the community that had been rooting for the company to build a massive frontier model.
The joke therefore lands unusually close to reality: Le Chaton Fat never existed, but a few months later Mistral is releasing an actual one-trillion-parameter open-weight flagship with a fat-cat codename.
A trillion parameters, but only 49 billion active
ML4 uses a sparse architecture: although it contains one trillion parameters in total, only 49 billion are active during inference. The design continues a broader trend among very large open models toward increasing total model capacity while activating only a fraction of the network at a time.
Mistral is also emphasizing how little hardware it says was required to train the system relative to the biggest frontier-model projects. Four thousand Blackwell GPUs is substantial in absolute terms, but the company argues it is modest against the resources available to larger U.S. labs. Exact efficiency comparisons are difficult because competitors do not publish complete training-compute data on a uniform basis. For comparison, Mistral said its previous Large 3 model — a 675-billion-parameter mixture-of-experts model with 41 billion active parameters — was trained on 3,000 Nvidia H200 GPUs.
The new model accepts multimodal inputs but still produces text output, Lample confirmed during an interview with VentureBeat. Mistral is targeting use cases including software engineering, cyber defense, financial analysis, satellite and aerial imagery, technical drawings and chip design.
Cybersecurity is the most strategically charged of those categories. Mistral argues that enterprises cannot depend entirely on closed providers whose safety systems may refuse dual-use but legitimate defensive requests. Its pitch is that an open-weight model gives security teams greater control over code scanning, defensive testing and other high-volume security workflows without depending on a provider’s changing moderation policy.
The benchmark picture is strong — but more complicated than Mistral’s charts suggest
Mistral supplied preliminary results showing ML4 at 62% on DeepSWE v1.1, a long-horizon software-engineering benchmark. Its chart lists Reflection AI’s new Beam model at 44%, Qwen 3.8 Max at 51%, DeepSeek V4 Pro 0813 at 57% and GLM-5.3 at 61%.
Those competitor figures are broadly traceable to public sources, but benchmark configuration matters.
Reflection’s own Beam announcement just yesterday reports 44.4% for Beam, 51% for Qwen 3.8 Max and 61% for GLM-5.3 under its comparison setup.
Artificial Analysis separately reports a 57% DeepSWE result for DeepSeek V4 Pro 0813 when paired with the Codex agent harness.
But the live DeepSWE leaderboard tells a different story when it selects the best published configuration for each model: it currently puts GLM-5.3 at about 69% and Kimi K3 at about 69%, with GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 around 74%.
In other words, ML4’s 62% preview result looks competitive, particularly against Western open-weight models such as Beam, but it does not establish an outright coding lead across every available model-and-agent configuration.
The legal benchmark provides a cleaner external cross-check. Mistral’s chart gives ML4 a 15% task-pass rate on Harvey’s Legal Agent Benchmark. Vals.ai’s public leaderboard currently reports Kimi K3 at 12.92%, MiMo V2.6 Pro at 10.83% and GLM-5.3 at 8.33%, matching the relevant rounded competitor figures in Mistral’s comparison. If ML4’s 15% was produced under the same methodology, it would put the model ahead of those open-weight rivals, although Vals currently has several proprietary models above that mark.
Mistral also reports 67% on Finch, an enterprise finance-and-accounting workflow benchmark, tied with DeepSeek V4 Pro 0813 in its supplied chart and ahead of GLM-5.3 at 65%. Finch itself is a public benchmark accepted to ACL 2026; its 384 tasks span long-running spreadsheet, document, search, modeling and reporting workflows drawn from messy enterprise-style data. But VentureBeat could not independently locate published Finch results for the exact newer-model scores shown in Mistral’s chart.
The same caveat applies to Mistral’s visual-grounding claims. Its materials show ML4 scoring 42% on Dense200 and 73% on DIOR-RSVG, ahead of most of the general-purpose models in its comparison. Dense200 and DIOR-RSVG are published visual-grounding benchmarks, but the exact GPT-6 Astra, Kimi K3 and DeepSeek results in Mistral’s slides do not appear in the public benchmark sources VentureBeat reviewed.
That distinction matters because Mistral’s largest strategic claim is not that ML4 beats every closed model overall, but that it is the strongest open-weight model developed outside China and competitive with the best Chinese open systems. As of press time, ML4 does not yet appear in Artificial Analysis’ public evaluations or the DeepSWE leaderboard, so Mistral’s ranking claim remains provisional until outsiders can test the final model and released weights.
Mistral’s bigger bet: open weights plus a sovereign enterprise stack
ML4 arrives at a pivotal moment for Mistral. Founded in 2023 by former DeepMind researcher Arthur Mensch and former Meta researchers Guillaume Lample and Timothée Lacroix, the company quickly became one of Europe’s most prominent challengers to U.S. and Chinese foundation-model labs.
Its first Mistral 7B model arrived in September 2023, followed by the sparse Mixtral 8x7B later that year, helping establish the company’s strategy around efficient, openly available model weights.
The company has since expanded well beyond model weights. Its stack now includes developer and enterprise products, customization services, inference infrastructure and Mistral Compute — part of a broader effort to sell enterprises not merely access to a model, but control over the infrastructure, deployment and engineering around it. Mistral’s product and news archive documents that expansion.
In September, Mistral announced a €3 billion Series D at a post-money valuation above €21 billion, a round the company described as the largest equity fundraising ever completed by a European technology company. Reuters reported the valuation at roughly $24 billion. Mistral says it now supports more than 125 global enterprises, including Airbus, ASML and HSBC.
That full-stack strategy also explains why Mistral is comfortable releasing powerful weights instead of keeping its flagship entirely behind an API. Lample told VentureBeat that customers increasingly need more than model access: deployment, infrastructure, customization and engineering support around increasingly complex AI workflows all form part of the business. Mistral is effectively betting that model weights will become more commoditized while the higher-value enterprise business shifts toward the systems built around them.
For enterprises, the practical question is whether ML4’s final checkpoint holds up once independent evaluators can test it — and whether Mistral can turn its sovereignty pitch into measurable advantages in cost, control and customization. The company is giving itself three weeks before the weights land publicly. After Oct. 27, the “Le Chonk” nickname will matter less than what developers can reproduce on their own hardware.