Mistral Large 4 Preview: A Trillion Parameters and a Company Report Card

Mistral announced a public preview of Mistral Large 4 on October 6, a natively multimodal model that the company describes as 1 trillion parameters with 49 billion active. Its own phrasing: "Unofficially ML4, very officially: le Chonk." Open weights are promised by the end of the month.

The announcement is a long list of wins. Most of them are reported by the company that won them.

What Mistral says it built

According to Mistral, ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe. The training data was multilingual, spanning more than 160 languages, including every official language of the European Union. The company says the model uses the same training, customization and RL environment it sells to customers through Mistral Forge. It calls ML4 the first milestone on the roadmap funded by its €3 billion Series D.

The specifications already disagree with each other. Mistral's blog post says 1 trillion parameters. Model documentation cited by The Decoder lists a fine-grained mixture-of-experts architecture with 1.05 trillion total parameters and a 1.6 billion parameter vision encoder. The cited material does not explain the difference between the two total counts.

The context window is a bigger gap. OpenRouter lists 512K tokens (524,288). The Decoder, citing the same model documentation, says one million. Which is correct is an open question. On the Mistral API, the model supports only two reasoning levels: "none" and "high," per Simon Willison.

The scorecard

Mistral's own figures, as reported in its announcement:

• Cyber Index: top five globally on the Artificial Analysis Cyber Index, and ahead of open-weight models developed outside China "by a wide margin."
• Vulnerability test: 82 percent on a test that asks a model to reproduce a real open-source vulnerability and then patch it. Mistral says that is the highest of any model, and that Claude Opus 5.5 and GPT-6 Astra scored near zero because of safety refusals.
• Cybench: 93 percent of the 40 exercises.
• Coding: 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 28.3 percent on Terminal-Bench 4, for a Coding Agent Index of 49.8 percent, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
• Work tasks: 59.9 percent on AutomationBench (657 business workflows) and 1,393 Elo on AA-Briefcase.
• Vision: 42 percent on Dense 200 against 41 percent for GPT-6 Astra.
• Security: 93.3 percent of attacks resisted on Lakera's B3 benchmark, and 1.691 out of 2 on KORA.

Note what the 82 percent measures. A score near zero for rivals that refuse the task tells you about their policies as much as their skills. That is a legitimate finding, but it is a different one from "better at finding bugs." The one-percentage-point gap on Dense 200 does not, by itself, establish a meaningful advantage; no variance information appears in the figures reported.

Now a broader independent index, and a human evaluation that Mistral also cites. The Artificial Analysis Intelligence Index, which aggregates ten benchmarks, puts ML4 at 38 points, according to The Decoder. That is up from 9 for Mistral Large 3 and 14 for Mistral Medium 3.5. It is still behind Claude Opus 5.5 Max at 58. In a blind human evaluation by Surge AI, ML4 Preview placed second of five models on coding quality at 3.74, behind Claude Opus 5 at 4.22. Second of five is respectable. It is not a leapfrog, whatever the headline framing, and TechCrunch's phrase was that the model is "aiming to leapfrog" American and Chinese rivals. Aiming is a verb with no eval attached.

The Decoder also reports that ML4 trails GLM-5.3 slightly on automated business workflows. Mistral's account is that the model was preferred in CAD and STEM and performed on par or close to GLM-5.3 in finance and coding.

What it means

The strategic pitch is not subtle. Mistral plans a European deployment that it operates end to end, independently of other digital service providers and under European law. In May, CEO Arthur Mensch told a French parliamentary commission that Europe risks becoming dependent on US models for cybersecurity. He said the French military's codebases shouldn't be scanned by Anthropic's Mythos. A model with strong reported security scores and expanded cyber capabilities fits that argument. The Decoder reports that Mistral renamed its chatbot Le Chat to Vibe in May.

That is also the risk. Until the weights are out, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get the same model "with reduced moderation and expanded cyber capabilities." How the model tells legitimate vulnerability research from preparation for an attack is an open question. The announcement, as reported, does not answer it.

What to watch

• The weights. They are due by the end of October. The license terms are unknown.
• The spec sheet. A final, official context window size, and whether the parameter count is 1 trillion or 1.05 trillion.
• The bill. The Decoder reports preview pricing of $0.68 per million input tokens and $2.09 per million output tokens, with cached inputs at $0.07; it also cites documentation listing double those rates, $1.36, $4.18 and $0.14. Which tier survives the preview is unknown.
• The independent runs. Third-party evaluators have already been cited (vals.ai on legal and financial tasks), but the cyber claims deserve repeat runs once the weights are public.
• The power supply. According to The Decoder, Mistral took out an $830 million loan in March for a datacenter near Paris and plans 200 megawatts of European compute capacity by the end of 2027.

The company behind Vibe is now asking for scrutiny by benchmark. Fine. Let's give it some.