Mistral launches Large 4 preview, a 1-trillion-parameter open-weight model

October 6, 2026, 14:00Source: Mistral AI

Mistral AI has launched a public preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters, 49 billion of them active. The company calls it its largest and most capable model to date and says the weights will be released by the end of the month.

Key points

Mistral Large 4 is a natively multimodal model with 1 trillion parameters and 49 billion active parameters, in public preview on Mistral Studio.
Mistral says the weights will be released by the end of the month.
The model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.
Mistral reports 82% on a vulnerability reproduce-and-patch test and 93% on Cybench, and says some closed models score near zero by refusing the task.
Coding results include 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.

What Mistral is releasing

Mistral Large 4, which the company informally calls ML4 and "le Chonk", is available now as a preview API on Mistral Studio. According to Mistral, it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company's own datacenters in Europe, and the preview is served from the same infrastructure.

Mistral says the model is competitive with the strongest open-source models globally and significantly outperforms any open-weight model developed in the US or Europe. On enterprise workloads such as cybersecurity, finance and law, the company says it is state-of-the-art among open models. In visual grounding, it claims the model surpasses even frontier closed models.

A significant share of the training data was multilingual, covering more than 160 languages, including every official language of the European Union. Mistral also says the model was trained with the same training, customization and RL environment it offers customers through Mistral Forge.

Cybersecurity is the headline use case

Mistral puts cybersecurity at the centre of the launch. On the Artificial Analysis Cyber Index, an independent evaluation of how well models find and fix security flaws in real software, ML4 ranks among the top five models globally, according to the company. It says the model leads open-weight models developed outside China by a wide margin.

On an index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82%, which Mistral says is the highest of any model.
It solves 93% of the challenges in Cybench, a set of 40 exercises drawn from security competitions.

Mistral argues that part of this lead comes from refusals. It says several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they decline the task. The company contends that defenders often need to prove a flaw is real before fixing it, work that safety filters in closed models can block. Open weights and self-deployment, it says, let organizations run security work under their own policies, on private cloud or on-premise.

Before the weights are published, Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities. These groups access the same model with reduced moderation and expanded cyber capabilities.

Coding, agents and multimodal results

For 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% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, according to the company. In a blind human evaluation run with Surge AI, professional annotators rated ML4 Preview second of five models at 3.74, ahead of Kimi K3 (3.59), GLM-5.3 (3.60) and GLM-5.2 (3.40), and behind only Claude Opus 5 (4.22).

On AutomationBench, which covers 657 business workflows across apps such as Gmail, Google Sheets, Slack and Salesforce, ML4 scores 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro. On AA-Briefcase, which evaluates long-horizon knowledge work, it reaches 1,393 Elo, ahead of DeepSeek V4 Pro.

Mistral describes the model's image understanding as a step change. It points to work on complex documents, charts, engineering drawings and gigapixel satellite imagery. On the Dense 200 visual grounding benchmark, it says ML4 edges GPT-6-Astra, 42% to 41%.

Science and math

Mistral says ML4 is state of the art on SciCode-Verified among open-weight models. It adds that the model can generate a full Hartree, Fock simulation in one shot, and that in human evaluations it reasons more precisely and with more structure than GLM-5.3 on math tasks.

What comes next

The preview is the first step. Mistral says it will share more details on the model architecture, additional benchmarks and its post-training methodology as it works toward the weight release. It also says ML4 will be the foundation for a new generation of specialized and optimized Mistral models.

For organizations that want strong capability without depending on a closed provider, the pitch is control. Mistral says the model will be available across multiple regions worldwide, including a European deployment that it operates end-to-end, independently of other digital service providers and under European law. The company has not detailed pricing in the announcement, and the benchmark results are its own reporting until the weights allow outside verification.