Mistral has announced a public preview of Mistral Large 4 (ML4), described as its largest and most capable open-weight model. The natively multimodal system contains one trillion parameters, with 52 billion active during operation. The preview is available through Mistral Studio, while model weights are expected by the end of October 2026.
ML4 is designed for coding, cybersecurity, agentic workflows, multimodal reasoning, science, finance, and legal work. Mistral reports strong benchmark results, including leading open-weight performance in cybersecurity, a 93% score on Cybench, and 82% on a vulnerability reproduction and patching test. It also claims competitive coding and terminal-workflow results, ranking second in a blind human coding evaluation. On business automation, ML4 reportedly outperforms several competing models across applications such as email, spreadsheets, collaboration tools, and customer-management systems.
The model’s multimodal capabilities include analyzing complex documents, charts, technical drawings, satellite imagery, and dense visual scenes. Mistral says ML4 surpasses GPT-6 Astra on its Dense 200 visual-grounding benchmark. It also supports scientific programming, mathematical reasoning, simulations, and technical research workflows, while showing strong results in legal and financial document and spreadsheet tasks.
Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, ML4 emphasizes AI sovereignty, self-deployment, and customer control. It supports more than 160 languages and will be available across multiple regions, including a European deployment operated under European law. Mistral is continuing red-team testing and plans to publish further architectural and training details.
Member takes
Why members shared this — posted without names.
1T and down to 49k is insane. still too large for my computer but could be sustainable
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Manipulation Score: LowHow much this article uses emotional framing vs. straight reporting
Loaded Language3525%
Absolutist2215%
Threat/Panic1825%
Us vs Them2515%
Engagement Bait3220%
Manipulation score look right?
Factual Claims (10)
mediumtechnology
Mistral is launching a public preview of Mistral Large 4, with model weights planned for release by the end of October 2026.
hightechnology
ML4 is a 1 trillion-parameter natively multimodal model with 52 billion active parameters.
mediumtechnology
ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.
hightechnology
ML4 ranks among the top five models globally on the Artificial Analysis Cyber Index and scores 82% on a vulnerability reproduction and patching test.
hightechnology
ML4 solves 93% of the challenges in Cybench, a set of 40 cybersecurity exercises.
mediumtechnology
ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.
mediumtechnology
ML4 ranked second of five models in a blind human coding evaluation conducted with Surge AI, scoring 3.74 out of 5.
mediumtechnology
ML4 scores 59.9% on AutomationBench and reaches 1,393 Elo on AA-Briefcase.
hightechnology
ML4 surpasses GPT-6-Astra on the Dense 200 visual-grounding benchmark, scoring 42% versus 41%.
mediumtechnology
A significant share of ML4's training data was multilingual and covered more than 160 languages, including every official language of the European Union.
Comments
Login to view and post comments.