Mistral unveils trillion-parameter Le Chonk

A very large cat enters the frontier race

Mistral’s new flagship model has an intentionally unserious nickname and a very serious strategic purpose. Large 4, or “Le Chonk,” was unveiled on October 6 as a 1-trillion-parameter, natively multimodal model with 49 billion active parameters, making it the Paris-based company’s largest and most capable model to date [1]. Mistral’s documentation lists the model more precisely at 1.05 trillion total parameters, 49 billion active parameters and a 1.6-billion-parameter vision encoder, with a 1-million-token context window [2].

That scale matters because Mistral is trying to reclaim a place at the frontier after months in which open-weight momentum appeared to favor Chinese labs, while the strongest closed systems remained concentrated in the United States. Mistral says ML4 is already competitive with the strongest open-source models globally and “significantly” ahead of any open-weight model developed in the US or Europe [1]. Reporting from TechCrunch framed the launch as an attempt to offer a European alternative to both closed American models and increasingly strong Chinese open-weight systems [4].

The launch is not a full weights release on day one. Mistral opened a public preview through its API and Mistral Studio, while saying the weights will drop by the end of the month [1]. VentureBeat reported that the target date for publishing the weights is October 27, after a roughly three-week testing period involving developers, cybersecurity leaders and government authorities [3]. TechCrunch likewise reported that the model is currently accessible through a moderated public endpoint, with weights expected after safety testing [4].

What Mistral is claiming

Mistral is emphasizing three dimensions: size, specialist performance and European control. On raw architecture, Le Chonk is a Mixture-of-Experts model, meaning only part of the full parameter set is active for a given request; that helps explain how a trillion-parameter model can be served with 49 billion active parameters [2]. On infrastructure, Mistral says the model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and that the public preview is served on the same infrastructure [1]. VentureBeat, citing materials it reviewed, reported the training took roughly two months on about 4,000 Nvidia Grace Blackwell GPUs in Mistral’s European data centers [3].

The company is also putting forward concrete benchmark claims. In cybersecurity, Mistral says ML4 ranks among the top five models globally on the Artificial Analysis Cyber Index and scores 82% on one vulnerability reproduction-and-patching test, which it describes as the highest score of any model on that test [1]. Mistral also says the model solves 93% of Cybench challenges, a set of 40 security competition exercises [1]. In coding, it reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index of 49.8% [1].

The benchmark story is not just a victory lap. Le Monde noted that Mistral’s broader independent ranking position had slipped before this release, and that some of the new performance claims still needed confirmation in regularly updated third-party rankings [7]. At the same time, the company says specific legal and financial evaluations were run through third-party evaluators, and it claims ML4 exceeded GPT-6-Astra on representative tasks in both areas [1]. In other words, the credible reading is that Mistral has published a strong set of preliminary and partner-evaluated results, while the open-model leaderboard story will continue to evolve after the weights are released.

Open weights, but not without guardrails

The “open-weight” label is central to the launch, but the timing is important. ML4 is being positioned as an open-weight model, yet the weights were not immediately downloadable at announcement time [4]. Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities before releasing the weights, and that those partners will test the same model with reduced moderation and expanded cyber capabilities [1].

That staged approach reflects the core tension around capable open models. Once weights are available, customers can inspect, customize and deploy the system on their own infrastructure; they also gain insulation from sudden provider access cuts. Mistral explicitly argues that this matters in cybersecurity, where refusals by a hosted provider can interfere with legitimate vulnerability research or incident response [1]. Axios reported that Mistral is initially making Le Chonk available through a moderated API while sharing a less restricted version with select partners, with weights planned for October 27 after further reinforcement learning and safety testing [5].

But openness also reduces the developer’s control over downstream use. Axios noted that once weights are public, users can modify the model and attempt to remove safeguards, a concern that becomes sharper when the same model is strong at cybersecurity tasks [5]. That is why the safety section of Mistral’s release matters: the company says ML4 resists 93.3% of attacks on Lakera’s public B3 AI Security Benchmark and reports its highest measured responsible-engagement score among open-source models on KORA Benchmark [1].

Europe’s sovereignty argument

Le Chonk is also a geopolitical product. Mistral says the model will be available across multiple regions, including a European deployment operated end-to-end by Mistral under European law [1]. The company also highlights multilingual training data spanning more than 160 languages, including every official language of the European Union [1]. Anadolu similarly reported that the model covers coding, cybersecurity, finance, legal tasks and image understanding, and that Mistral is stressing deployment control for organizations using open-weight models [6].

That sovereignty pitch has commercial substance. Enterprises and governments do not only buy model quality; they buy predictable access, data-control promises, regulatory posture and the ability to deploy in private clouds or on-premise environments. Mistral is arguing that ML4’s open-weight path gives security-sensitive customers a way to keep advanced AI capabilities under their own policies [1]. Le Monde reported that the company is counting on expanded European data-center capacity over the next six months and sees Large 4 as a foundation for future Large 5 and Large 6 improvements [7].

The cost side is the obvious caveat. A trillion-parameter inventory is impressive, but even with MoE efficiency, this is not a toy model that most teams will casually self-host. Mistral’s own documentation lists API pricing for Large 4 and a 1-million-token context window, signaling that inference economics will be part of the adoption calculus [2]. The Dark Souls joke almost writes itself: open weights may give you the item, but carrying Le Chonk can still feel like maxing out your equipment load.

Why it matters

The significance of Le Chonk is not that Europe has suddenly surpassed every rival. TechCrunch quoted Mistral’s Pierre Stock as saying the company hopes ML4 will be best in class among open-weight models, especially outside China, while acknowledging benchmark results were still pending at launch [4]. Axios also reported Stock’s admission that Mistral has not yet caught the leading closed models [5]. That restraint is important: Le Chonk is a major European bid, not definitive proof that the frontier has moved to Paris.

Still, the release changes the conversation. Mistral now has a trillion-parameter flagship with public benchmark claims, a near-term open-weight release plan and a sovereignty story that fits the needs of European governments and regulated industries. If the October 27 weights arrive as promised, the next test will be brutally practical: how well the model performs outside Mistral’s demos, how much it costs to run, how safely it can be adapted, and whether developers find that this particular Chonk is worth the carry weight.

Sources from the last 72 hours

1. [1]Introducing Mistral Large 4 | MistralOct 6, 2026, 2:00 AM
2. [2]Mistral Large 4 - Mistral AI | Mistral DocsOct 6, 2026, 2:00 AM
3. [3]Mistral debuts Large 4 ‘Le Chonk', a 1-trillion parameter text output model with high benchmarks planned for open weights releaseOct 6, 2026, 3:00 PM
4. [4]Mistral’s new 1T model aims to leapfrog closed and open rivalsOct 6, 2026, 4:33 PM
5. [5]Western AI labs challenge China's open-model leadOct 6, 2026, 2:00 AM
6. [6]French AI firm Mistral unveils AI model it says rivals strongest Chinese open systemsOct 6, 2026, 2:00 AM
7. [7]Mistral AI unveils new AI model aimed at 'narrowing the gap' with top Chinese competitorsOct 6, 2026, 2:48 PM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.