Mistral drops 1T-parameter 'Le Chonk' open model

Europe just got its most powerful open-weight model yet — and it has a ridiculous name.

Mistral AI launched Mistral Large 4 on October 6, 2026 (UTC), a 1.05 trillion-parameter Mixture-of-Experts model nicknamed "Le Chonk." The model is available now in public preview through Mistral's API, with open-weight downloads scheduled for October 27, 2026 (UTC). If the release holds, it will be the largest open-weight model from a Western company — and a direct challenge to the Chinese open models that have dominated the open ecosystem all year.

The specs

Mistral Large 4 is a sparse MoE model with 1.05 trillion total parameters and 49 billion active parameters per token. It includes a 1.6 billion-parameter vision encoder for native multimodal input, supports a 1 million token context window, and is fluent in over 160 languages. The model was trained over roughly two months on nearly 4,000 Nvidia Grace Blackwell GPUs, built end-to-end in Europe, and can be deployed from Mistral's own cloud infrastructure based in the EU.

Pricing is aggressive: $0.68 per million input tokens, $0.07 per million cached input tokens, and $2.09 per million output tokens. For comparison, GPT-6 Astra costs $10/$50 and Claude Opus 5.5 costs $4/$20. Mistral is positioning Large 4 as a cost-efficient frontier alternative, not a direct match for the absolute top closed models.

Sources: Mistral AI, OpenAI, Anthropic, model documentation

What Mistral is claiming

Mistral says Large 4 is the best open-weight model from the United States or Europe on aggregated benchmarks. The company specifically calls out state-of-the-art results in cyber defense, manufacturing, and finance workloads, and claims the model outperforms closed frontier models on visual grounding tasks.

Independent third-party data tells a more nuanced story. On the Artificial Analysis Intelligence Index, which aggregates ten benchmarks across domains, Mistral Large 4 scores 38 — making it the top-ranked non-Chinese open model, but still well behind Kimi K3 at 58. Mistral itself concedes that "frontier open models like Kimi K3 remain ahead on raw capability" and stakes its claim on efficiency instead: scores comparable to GLM-5.2 on advanced reasoning while using a fraction of the active parameters.

The model is currently in public preview, and Mistral is conducting private security reviews with cybersecurity partners before the weight release. The reinforcement learning run behind the preview is still in progress, meaning the final open-weight checkpoint may differ from what's available through the API today.

Why it matters

The open-weight landscape has been dominated by Chinese models all year. Kimi K3, DeepSeek V4, and GLM-5.3 have set the pace, while Western companies have largely kept their best models behind APIs. Mistral Large 4 changes that calculus. At 1.05 trillion parameters with 49 billion active, it's the first Western open model that can credibly compete on raw scale with the Chinese frontier open models.

The timing is significant. Reflection AI announced its 501B parameter Beam model just one day earlier, on October 5, 2026 (UTC), with weights promised for later this month. Now Mistral is following with a model more than twice the size. The open-weight space is suddenly crowded — and competitive — in a way it hasn't been since the original Llama 3 release.

For European sovereignty, this is the model that policymakers have been asking for. Built entirely in Europe, deployable from European infrastructure, with open weights that any organization can audit and run locally. Mistral has been the standard-bearer for European AI, and Large 4 is its most ambitious release yet. The question is whether "built in Europe" is enough to compete with models that have access to vastly more training compute and data.

Critical lens

There are reasons for caution. Mistral's benchmark claims come primarily from internal evaluations. The Artificial Analysis score of 38 is respectable, but the 20-point gap to Kimi K3 is substantial — roughly the difference between a top-tier model and a mid-tier one. If the final open-weight release doesn't close that gap, Large 4 may end up as a "best in the West" consolation prize rather than a true frontier competitor.

The security review before weight release is also telling. Mistral is clearly concerned about what a 1 trillion parameter open model could do in the wrong hands — particularly given its claimed strength in cyber defense (and, by extension, cyber offense). The October 27 release date gives the company three weeks to conduct red-teaming, but whether that's enough for a model of this size is an open question.

The pricing is interesting too. At $0.68/$2.09, Mistral is undercutting even mid-tier closed models. That suggests the company is betting on volume over margin — and that it believes open-weight models will win on cost, not absolute capability. That's a reasonable strategy, but it depends on the model being "good enough" for most enterprise use cases. The 38-point Intelligence Index score raises questions about whether it clears that bar.

What to watch

• October 27 weight release: The actual checkpoint will tell us whether the preview performance holds. Watch for independent benchmark results within 48 hours of release.
• Kimi K3 response: Will Moonshot AI update Kimi K3 or release a new model to maintain its lead? The open-model race just got real.
• Enterprise adoption: The real test is whether companies choose Mistral Large 4 over closed APIs. Watch for major enterprise announcements in the November-December timeframe.
• Security incidents: A 1T open model with cyber capabilities is a significant dual-use risk. Any reported misuse in the first month after release will shape the conversation around open-weight frontier models.
• Reflection Beam comparison: Both models promised open weights within weeks of each other. A head-to-head comparison once both are available will tell us which approach — 501B/23B vs 1.05T/49B — delivers better real-world performance.

Mistral Large 4 isn't going to dethrone GPT-6 or Claude Fable. But it doesn't need to. What it does is give the open-weight ecosystem a Western flagship at a scale that can compete — on price, on sovereignty, and on the fundamental promise that the most powerful AI shouldn't be locked behind a handful of API keys. Whether "Le Chonk" can deliver on that promise is something we'll know for sure on October 27.