Mistral Large 4: a trillion parameters, open weights, and cyber front and center

A trillion parameters, European style

Mistral is launching the public preview of Mistral Large 4, unofficially nicknamed 'the Chonk.' It's the biggest model in the company's history: 1 trillion total parameters, 52 billion active at each inference, and native multimodality.

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs installed in Mistral's own datacenters in Europe. The preview runs on that same infrastructure. Mistral also promises a fully European deployment, operated end to end, independent of other digital service providers and subject to European law.

The weights will be released by the end of the month, along with architecture details, additional benchmarks, and post-training methodology. In the meantime, the model is undergoing red-teaming with cybersecurity actors, select partners, and state authorities, who have access to a version with reduced moderation and extended cyber capabilities.

Cybersecurity as a selling point

This is the most emphasized angle of the announcement. On the Artificial Analysis Cyber Index, an independent evaluation of the ability to find and fix flaws in real software, ML4 ranks in the global top 5 and clearly outpaces open-weight models developed outside China.

On one of the index's tests, which consists of reproducing a real vulnerability in open source software and then patching it, ML4 reaches 82%, the best score of any model. Claude Opus 5.5 and GPT-6 Astra score near zero there, not out of incapacity, but because they refuse the task.

Mistral draws a substantive argument from this: defending software often starts with proving a flaw exists, and closed models' filters block exactly that kind of work, while attackers simply jailbreak those same models. Losing access to a capability mid-incident is itself a risk. Hence the promise of a model deployable in private cloud or on-premise, under the organization's own policies.

Code, agents, and office work

In agentic coding, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its Coding Agent Index of 49.8% puts it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind human evaluation run with Surge AI, it finishes second out of five (3.74/5), ahead of GLM-5.3, Kimi K3, and GLM-5.2, but clearly behind Claude Opus 5 (4.22).

On AutomationBench, 657 business workflows involving Gmail, Google Sheets, Slack, or Salesforce, it reaches 59.9%. On AA-Briefcase, which evaluates long-horizon knowledge work, it scores 1,393 Elo, ahead of DeepSeek V4 Pro.

On the regulated-industries side, vals.ai judges it superior to GPT-6-Astra on representative legal and financial tasks, and it beats all open source models on Harvey's Legal Agent benchmark.

Vision, science, and math

Mistral presents ML4 as a leap forward in image understanding: complex documents, charts, natural scenes, gigapixel satellite imagery, technical drawings. The model combines visual grounding with agentic capabilities to zoom, inspect, and verify until it reaches an accurate answer. On Dense 200, it edges out GPT-6-Astra, narrowly (42% vs. 41%).

In science, it's announced as state of the art among open-weight models on SciCode-Verified, and capable of generating a complete Hartree–Fock simulation in one pass, a multi-step chemistry task. In an internal human evaluation, it's preferred over GLM-5.3 in CAD and STEM, and is roughly on par or just slightly behind in finance and code.

Model safety: the embraced paradox

ML4 resists 93.3% of attacks in Lakera's B3 benchmark and saturates internal tests of robustness against indirect prompt injections. On the KORA Benchmark, it scores 1.691 out of 2, the best score Mistral measured among open source models.

Notable detail: despite its cyber performance, its average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is higher than that of every other open source model. Mistral is thus claiming both a model that accepts defensive work and refuses malicious requests more often, a balance that only real-world use will be able to confirm.

Industrial-scale RL, and what's next

Post-training relies heavily on reinforcement learning, via a library where environments (chat, scientific problems, alignment, factuality, long-horizon tool use) are combined within a single run. An autoscaling fleet of actors generates tens of thousands of rollouts in parallel, with asynchronous training and budgets of millions of tokens per trajectory.

On roughly 3,000 GPUs, a run produces nearly 33 billion tokens per day, of which 16 billion are usable for training. This run is still ongoing and, according to Mistral, shows no saturation: the company promises rapid improvements in the coming weeks.

ML4 is presented as the first milestone of the €3 billion Series D, the largest funding round ever raised by a European tech company. It will also serve as the foundation for a new generation of specialized Mistral models, trained using the same environment offered to customers via Mistral Forge.