Mistral Named the Trillion-Param Model Le Chonk

Mistral could have called it Large 4 and gone home. It went with le Chonk.

The public preview opened 6 October 2026 on Mistral Studio. The model is a natively multimodal mixture of experts: about 1 trillion parameters, 49 billion active, a 1.6-billion-parameter vision encoder, and a claimed 1-million-token context on Mistral's own card. Vercel's gateway lists 524,288. Believe the smaller number until you measure. Weights are promised by the end of the month. Until then, Mistral is red-teaming with cybersecurity teams, partners, and state authorities on a less-moderated build.

Image: Mistral, Introducing Mistral Large 4, 6 October 2026.

The numbers they want quoted

DeepSWE v1.1: 61.7%. SWE-Atlas-QnA: 59.4%. Terminal-Bench 4: 28.3%. Combined Coding Agent Index: 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max on Mistral's chart. Surge AI blind coding quality: 3.74 / 5, second of five, behind Claude Opus 5 at 4.22. AutomationBench: 59.9% on 657 business workflows. AA-Briefcase: 1,393 Elo. Visual grounding, Dense 200: 42% vs GPT-6 Astra at 41%. SciCode-Verified: state of the art among open-weight, per Mistral. vals.ai legal and finance tasks: ahead of GPT-6 Astra in both, Mistral says. Harvey's Legal Agent: best among open-source on their telling.

Cyber is the wedge

On Artificial Analysis Cyber Index, Mistral says ML4 sits in the global top five and leads non-Chinese open-weight by a wide margin. One test — reproduce a real open-source vulnerability, then patch it — scores 82%, the highest of any model they list. Cybench: 93% of 40 CTF-style tasks. The comparison they lean on: Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduce-the-bug test because they refuse. Mistral's argument is that defenders need a model that will do the work closed labs block, and that open weights plus a European datacenter are how you keep that capability when a vendor changes its mind. Training ran from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European sites. More than 160 languages in the mix, including every EU official language. Series D was €3 billion.

Safety numbers sit next to the cyber brag, on purpose. Lakera B3: 93.3% attack resistance. KORA: 1.691 / 2. Average cyber-refusal on JailbreakBench, StrongREJECT, and AgentHarm is higher than other open-weight models, Mistral says, even while the cyber benches are strong. That is a marketing knot: capable and unwilling, except for the customers who get the reduced-moderation build. Ask which SKU you are buying.

Preview pricing on Mistral's standard tier, as aggregated from the model page the same day: $0.68 / $0.07 cached / $2.09 per million tokens. Some third-party listings doubled that. Check the live table before you forecast. The useful date is not today. It is the weight drop. Until the checkpoint is on a disk you control, le Chonk is an API with a nickname.

Sources

Mistral, Introducing Mistral Large 4, 6 October 2026. https://mistral.ai/news/mistral-large-4/

Carl Franzen, VentureBeat, 6 October 2026. https://venturebeat.com/technology/mistral-debuts-large-4-le-chonk-a-1-trillion-parameter-text-output-model-with-high-benchmarks-planned-for-open-weights-release