Mistral Puts Its Trillion-Parameter Large 4 Into Public Preview, With Weights Promised by Month’s End

Mistral AI launched a public preview of Mistral Large 4 on October 6, 2026. The company calls it its largest and most capable model so far: a natively multimodal model with 1 trillion parameters, 49 billion of which are active, according to the launch post. Developers can call the preview through the API on Mistral Studio now. The weights are not public yet. Mistral says it “will release the weights by the end of the month” and will share more on the architecture, extra benchmarks, and its post-training method at that point. Until then, the company is red-teaming the model with “cybersecurity leaders, vetted partners, and state authorities,” who get access with reduced moderation and expanded cyber capabilities.

The documentation page, also dated October 6, labels the model “Public Preview” and version v26.10. Its figures differ a little from the launch post. It lists 1.05 trillion total parameters, 52 billion active parameters, and a separate 1.6-billion-parameter vision encoder, and describes a “granular Mixture-of-Experts architecture” with a 1M-token context. The page shows input at $0.68 per million tokens, cached input at $0.07, and output at $2.09, each listed beside a higher rate of $1.36, $0.14, and $4.18.

Training it on European hardware

Mistral says ML4, the model’s informal short name, was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the preview runs on the same infrastructure. The company says it will offer the model in several regions worldwide, including a European deployment it operates end to end “independently of other digital service providers and under European law.” Its training data spans more than 160 languages.

The post also describes Mistral’s reinforcement learning setup. An autoscaling fleet of actors generates tens of thousands of rollouts in parallel while training runs asynchronously. Rollout budgets reach millions of tokens across several context compactions. At about 3,000 GPUs, Mistral says, one training run produces roughly 33 billion tokens a day, of which about 16 billion are trainable completion tokens after filtering and masking. The company adds that the RL run behind the preview “is still in flight” and shows no sign of saturating.

The cyber pitch, and its refusal numbers

Security is the use case Mistral pushes hardest. On the Artificial Analysis Cyber Index, the company says ML4 ranks in the top five models globally and leads open-weight models developed outside China. On one of the index’s tests, which asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82%, the highest of any model by Mistral’s account. It also solves 93% of Cybench, a set of 40 exercises taken from security competitions.

Mistral argues that several closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the reproduce-and-patch test because they refuse it. Its case is that defenders need a model that can prove a flaw is real and that cannot be cut off mid-incident. The company also reports that ML4’s average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is higher than that of every open-source model it compared. It says the model resists 93.3% of attacks on Lakera’s public B3 AI Security Benchmark.

Coding, agents and documents

On software engineering, Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its combined Coding Agent Index score of 49.8% puts it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind human evaluation run with Surge AI, annotators rated ML4 Preview second of five models at 3.74 on a five-point scale. That was ahead of GLM-5.3 at 3.60, Kimi K3 at 3.59, and GLM-5.2 at 3.40, and behind Claude Opus 5 at 4.22.

For agent work, the post cites 59.9% on AutomationBench, a set of 657 business workflows across apps such as Gmail, Google Sheets, Slack, and Salesforce, and 1,393 Elo on AA-Briefcase for long-horizon knowledge work. On visual grounding, Mistral says ML4 edges GPT-6-Astra on Dense 200, 42% to 41%.

Mistral either produced all of these numbers itself or selected them for its announcement. No model card or independent evaluation of the open weights exists yet, because the weights have not shipped.

Where the preview leaves developers

For now, ML4 is a hosted API model with an open-weights commitment and a date attached. Mistral calls it the first milestone funded by its €3 billion Series D and says it will be the base for a new generation of specialized Mistral models. Teams that need self-hosting, the main argument behind the cybersecurity pitch, will have to wait for the end-of-month release. That release should also show the license, the final parameter count, and whether the shipped checkpoint matches the preview.