Mistral previews Large 4, a 1.05T multimodal model

Mistral previewed Large 4 (Le Chonk) on October 6, 2026: a 49B-active, 1.05T open-weight multimodal MoE, with weights due by late October.

On October 6, 2026, Mistral published Mistral Large 4 (ML4), nicknamed Le Chonk, in public preview (v26.10): an open-weight, general-purpose, natively multimodal Mixture-of-Experts model with 49 billion active parameters, a 1.6 billion-parameter vision encoder, native image input, and a 1 million-token context. Sources disagree on total size—the most specific figure is 1.05 trillion parameters, while others round to 1 trillion—and on training scale: several reports say 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, while TechCrunch says 4,000 NVIDIA GPUs on Mistral’s own compute, which the company says is two to three times less than Chinese competitors. The preview API is live, including on Mistral Studio (TechCrunch describes a guarded endpoint), at $0.68 per million input tokens, with structured outputs, function calling, document Q&A, agents, and built-in tools, and only “none” and “high” reasoning levels; weights are still unreleased, expected by the end of October 2026 or in about three weeks after red-teaming and safety testing. Reported results include 38 on the Artificial Analysis Intelligence Index (up from Large 3’s 9, ahead of GLM-5.2 and just behind DeepSeek 4.1 Flash, a 552B model, behind Claude Opus 5.5 at 58), 82% on a vulnerability reproduction-and-patch test that Claude Opus 5.5 and GPT-6 Astra reportedly score near zero on by refusing, 93% on Cybench, a claimed top-five Cyber Index place, and 49.8% on the Coding Agent Index ahead of DeepSeek V4 Pro and Qwen3.8 Max, though TechCrunch says independent benchmarks are still pending. Funding accounts also differ—a €3 billion Series D said to fund training versus ASML- and Samsung-backed rounds, with TechCrunch citing a Samsung-led Series D at a €21 billion valuation (about $24.39 billion)—and stated focus areas include cybersecurity, finance, and chip design. A Hacker News thread had 46 points and four comments, where Simon Willison compared the model with Claude Opus 5.5, GPT-6.1 Sol, and Gemini 3.8 Flash on a whimsical SVG prompt and argued ordinary benchmarks are saturated.

• On October 6, 2026, Mistral put Mistral Large 4 (ML4, “Le Chonk”) in public preview (v26.10): a natively multimodal MoE with 49B active parameters, 1.05T total (some reports round to 1T), a 1.6B vision encoder, native image input, and a…
• The preview API is live (Mistral Studio; TechCrunch calls it a guarded endpoint) at $0.68 per million input tokens, with structured outputs, function calling, document Q&A, agents, built-in tools, and only “none” and “high” reasoning…
• Open weights are not released; sources say end of October 2026, by month’s end, or in about three weeks, after red-teaming and safety testing with security firms, partners, and governments.
• Training scale disagrees: 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters versus TechCrunch’s 4,000 NVIDIA GPUs on Mistral’s own compute, said to be two to three times less than Chinese rivals.
• Scores cited: 38 on the Artificial Analysis Intelligence Index (Large 3 was 9; ahead of GLM-5.2; just behind DeepSeek 4.1 Flash, 552B; behind Claude Opus 5.5 at 58), 82% on a vuln reproduce-and-patch test, 93% on Cybench, a claimed…
• Funding accounts disagree: a €3 billion Series D said to fund training versus ASML- and Samsung-backed rounds, with TechCrunch citing a Samsung-led Series D valuing Mistral at €21 billion (about $24.39 billion).
• A Hacker News thread showed 46 points and four comments; Simon Willison compared ML4 with Claude Opus 5.5, GPT-6.1 Sol, and Gemini 3.8 Flash on a whimsical SVG prompt.