Introducing Mistral Large 4

On 6 October 2026, Mistral opened a public preview of Mistral Large 4, which the company calls ML4 and, with some humour, "le Chonk". It is a natively multimodal model with 1 trillion parameters in total, of which 52 billion are active for any given request. Mistral describes it as its largest and most capable model so far. Developers can try the preview API on Mistral Studio today, and the company said the weights would follow by the end of October.

Trained and served in Europe

Mistral built ML4 from the ground up using 3,800 NVIDIA Grace Blackwell GPUs, housed in datacenters that it owns in Europe. The preview runs there too. The model will be offered in several regions around the world, including a European deployment that Mistral itself runs completely, with no other digital service provider involved and subject to European law. Much of the training data is multilingual, covering over 160 languages. That includes all the official languages of the EU.

During training, Mistral consulted large customers from banks and factories to drug makers, shipping lines and government agencies. The environment it used for training and reinforcement learning is the one it sells to customers under the Forge name.

A case built on cybersecurity

The strongest claim concerns security. The Artificial Analysis Cyber Index, run independently, measures how well a model can find and repair weaknesses in real software. Mistral reports that ML4 sits in the global top five there and is well ahead of any open-weight model made outside China. One test in the index asks the model to reproduce a known flaw in an open-source project and then patch it, and ML4 scored 82%, the best result of any model. It also cleared 93% of Cybench, a set of 40 exercises taken from security competitions.

Mistral notes that several closed models score close to zero on that same test, because they decline to do the work. Defenders, it argues, often need to demonstrate that a flaw is real before they can fix it, and a provider's refusals can stall research or incident response. A model with open weights can be run by the customer under its own rules, in a private cloud or on site. While the weights are still unreleased, Mistral is red-teaming ML4 with security specialists, vetted partners and state authorities, who get a version with lighter moderation and wider cyber abilities.

Coding, agents and office work

On software engineering, the company cites 61.7% for DeepSWE v1.1, 59.4% for SWE-Atlas-QnA and 28.3% for Terminal-Bench 4. Blended into a single Coding Agent Index, that gives 49.8%, which Mistral says beats DeepSeek V4 Pro 0813 and Qwen3.8 Max. A blind test with Surge AI had professional annotators grade the output of five models from 1 to 5. ML4 Preview came second with 3.74, behind Claude Opus 5 (4.22) and ahead of Kimi K3, GLM-5.3 and GLM-5.2.

For general agent work, AutomationBench tests 657 business workflows in apps like Gmail, Google Sheets, Slack and Salesforce, and ML4 reaches 59.9%. Mistral adds that vals.ai, a third-party evaluator, found it ahead of GPT-6-Astra on legal and finance tasks, and that it tops every open-source model on the Legal Agent benchmark from HarveyAI.

Vision, science and safety

The model reads images far better than its predecessors, according to Mistral, with uses in engineering, manufacturing and earth observation such as scanning gigapixel satellite images for disaster response or checking engineering drawings. On visual grounding it edges past GPT-6-Astra on Dense 200, 42% to 41%. In science, Mistral calls it state of the art among open-weight models on SciCode-Verified and says it can write a full Hartree–Fock simulation in one pass.

On safety, ML4 resists 93.3% of attacks on Lakera's public B3 AI Security Benchmark. Mistral also says its refusal rate on malicious cyber prompts is higher than that of any open-source model it compared.

The first milestone for the Series D

Mistral frames ML4 as the first milestone on the roadmap funded by its €3 billion Series D. The reinforcement-learning run behind the preview is still going, at about 3,000 GPUs, and the company expects large gains in the weeks and months ahead. It plans to publish more on architecture, benchmarks and post-training, and to use ML4 as the base for a new generation of specialised models.

All benchmark figures in this article are Mistral's own.