Mistral Large 4 "Le Chonk": 1T-Parameter Open Model Goes Public

Mistral Large 4 is here, and Mistral is calling it "Le Chonk." The French lab opened a public preview of the 1-trillion-parameter model on October 6, promised open weights by the end of the month, and independent tester Artificial Analysis already rates it the smartest model from outside the US and China, though still well behind the top closed models.

What actually happened

Mistral describes the model as a natively multimodal, hybrid instruct-and-reasoning model with more than 160 languages. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview runs on the same hardware.

It is also the first big release funded by Mistral's €3 billion Series D, which the company calls the largest equity round ever raised by a European tech company. The weights are not out yet. Mistral says it will share architecture details, more benchmarks and its post-training method as it works toward the release.

There is a sovereignty angle too. Mistral says the model will be served in multiple regions, including a European deployment that it runs end to end, independently of other digital service providers and under European law. For EU companies nervous about US clouds, that is a real selling point.

The Register's take is blunt and fair: on Mistral's own charts, Mistral Large 4 trades blows with top Chinese open models from Alibaba, Moonshot, Z.ai and DeepSeek. Independent testing puts it below OpenAI and Anthropic flagships, but well ahead of Thinking Machines Lab's Inkling, the strongest US open-weights model.

The benchmarks, company vs independent

Mistral's numbers: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4, and 49.8% on its combined Coding Agent Index. It also says the model resists 93.3% of attacks on Lakera's B3 security benchmark. These are self-reported.

Artificial Analysis, as reported by Unite.AI, gives the independent view. Beyond the Intelligence Index score of 38, the preview scored 50 on its Cyber Index and 82% on CyberGym-E2E, ahead of GPT-6 Luna (max) at 78% on that test. It measured output at 116.1 tokens per second and time to first token at 1.46 seconds through Mistral's API.

One thing to know about cost: the preview generated 200 million output tokens over the Intelligence Index run, against a class median of 81 million. Artificial Analysis puts the run at $1.13 per task, or $0.57 with the launch discount. Cheap per token can still mean wordy per answer.

The cyber pitch

Mistral is leaning hard into security. It claims leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on one of its reproduce-and-patch tests because they refuse the task. Its argument is that refusals can block legitimate vulnerability research, and that self-hosted open weights fix that.

Mistral also says the model's refusal rate on malicious cyber prompts is higher than all other open models it tested. So the pitch is "useful for defenders, stubborn with attackers." That is a company claim worth testing, not a guarantee.

What Polymarket thinks

Polymarket opened "Mistral Large 4: Text Arena Debut?" on October 7. At 06:10 UTC on October 8, the Polymarket arena market had 1430+ at 81.5%, 1450+ at 53%, 1470+ at 30.5% and 1490+ at 11.5%. It is a thin market at about $13,800 in volume.

It resolves on the model's Arena text score at noon ET on the day after it first appears on the leaderboard. A coin-flip on 1450+ says traders expect a solid but not top-tier debut.

What to do as a reader

If you build with open models, Mistral Large 4 is worth a test once the weights land, especially for multilingual or security work where you want to run models on your own hardware. Try the API preview now, and use the two-week discount while it lasts.

Do not take launch charts at face value. Compare Mistral's self-reported scores with Artificial Analysis and the Arena once it lists. For more on the open-model race, see our Kimi K2.6 and Kolibri open-weight coverage, and our October model odds roundup.