Mistral Large 4: 1-trillion-parameter model in preview, weights to follow

What this is about

Mistral AI has launched a public preview of Mistral Large 4 (internally "ML4", jokingly "le Chonk"). According to Mistral, it is the company's largest and most capable model to date: 1 trillion parameters in total, 52 billion of them active. Parameters are, roughly speaking, the dials learned during training. "Active" means that only a portion of them is used for each request (mixture of experts, MoE). The model is natively multimodal, meaning it processes text and images together from the ground up, and combines instruction following, reasoning (step-by-step thinking) and agentic capabilities in a single model.

The preview API is available through Mistral Studio, and the weights are due at the end of the month. Until then, Mistral is testing the model in real-world use with cybersecurity professionals, vetted partners and government agencies. They get the same model with reduced moderation and extended cyber capabilities. Pricing is $1.36 per million input tokens and $4.18 per million output tokens.

Training and European infrastructure

ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers in Europe, and the preview runs on the same infrastructure. The model is meant to be available in several regions worldwide, including a European variant that Mistral operates entirely itself, independent of other digital service providers and under European law. A large share of the training data was multilingual and covers more than 160 languages, including all official EU languages.

During training, Mistral says it worked with companies from finance, industry, logistics, pharma, science, shipping and the public sector. The training, customization and RL environment match what customers can use through Mistral Forge.

On post-training: Mistral relies heavily on reinforcement learning (RL), meaning learning from the outcomes of the model's own attempts. The library lets it combine tasks in a single run, from simple chat to scientific problems to safety, factuality and long tool-use workflows. At around 3,000 GPUs today, a training run generates about 33 billion tokens per day, of which roughly 16 billion are trainable response tokens. The RL run is still going. Mistral sees "no signs of saturation" and expects rapid improvements.

Performance according to Mistral

All figures below are vendor-reported.

Cybersecurity: On the Artificial Analysis Cyber Index, ML4 ranks among the top five models worldwide and leads clearly among open models outside China. In a sub-test that involves reproducing and patching a real vulnerability in open-source software, it scores 82 percent, the highest of any model. On Cybench (40 tasks from security competitions), it solves 93 percent. Mistral points out that Claude Opus 5.5 and GPT-6 Astra score close to zero on the same test because they refuse the task. At the same time, ML4 is said to refuse requests with malicious intent more often than other open models (measured with JailbreakBench, StrongREJECT and AgentHarm).

Coding: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4. The Coding Agent Index stands at 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind test with Surge AI, professionals rated outputs on a scale of 1 to 5:

Agents and knowledge work: On AutomationBench (657 business workflows in Gmail, Google Sheets, Slack and Salesforce), ML4 scores 59.9% and leads Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro. On AA-Briefcase, it reaches 1,393 Elo. On legal and finance tasks, it beats GPT-6 Astra according to third-party vals.ai, and on HarveyAI's Legal Agent Benchmark it beats all open-source models.

Images, science, math: On image understanding (documents, charts, satellite images, technical drawings), Mistral speaks of a quantum leap. On visual grounding (locating objects in an image), it narrowly beats GPT-6 Astra on Dense 200 (42% vs. 41%). On SciCode-Verified, it leads among open models. In human comparisons against GLM-5.3, ML4 was preferred on CAD and STEM, and was on par or slightly behind on finance and coding.

Safety: Against indirect prompt injection, Mistral considers its benchmarks saturated. On the Lakera B3 benchmark, ML4 resists 93.3% of attacks. On the KORA benchmark, it scores 1.691 out of a maximum of 2.

Our take

Mistral positions ML4 as a European, sovereign alternative: open weights, self-hosting and operation under European law. The cybersecurity argument is especially notable: defenders need models that don't refuse vulnerability research. It is also a delicate matter, because the extended version with reduced moderation goes only to vetted partners. How the public weights will be configured has not yet been decided.

A few caveats: almost all figures come from Mistral or from selected evaluators, and many are internal tests. On coding, ML4 trails the best model in the blind test (Claude Opus 5) by a wide margin, and its 28.3% on Terminal-Bench 4 is far from the scores of other recent models. Architecture details, further benchmarks and the license are still open and are due with the weights. At 1 trillion parameters, running it locally is realistic only with server hardware. Mistral points to a €3 billion Series D as funding for expanding its compute capacity.