Mistral Large 4 Release Notes

Mistral Large 4 (ML4), unofficially dubbed "le Chonk," is a 1 trillion-parameter natively multimodal model designed to push the frontier of open-weight performance. The model is currently available in public preview via the Mistral Studio API, with full weights scheduled for release by the end of October 2026.

Model Architecture and Infrastructure

ML4 is a Mixture-of-Experts (MoE) model featuring 1 trillion total parameters and 49 billion active parameters. It was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs within Mistral's own European datacenters.

To ensure AI sovereignty, the model is deployed on infrastructure operated end-to-end by Mistral under European law, independent of other digital service providers. The training data was extensively multilingual, covering over 160 languages, including all official European Union languages.

Cybersecurity and Safety Performance

ML4 is positioned as a state-of-the-art open-weight model for cybersecurity, specifically designed to avoid the "provider-level refusals" common in closed models that can hinder legitimate vulnerability research.

Cybersecurity Benchmarks

  • Artificial Analysis Cyber Index: ML4 ranks among the top five models globally and leads open-weight models developed outside China.
  • Vulnerability Reproduction: ML4 scored 82% on a test requiring the reproduction and patching of real open-source software vulnerabilities, outperforming closed models like Claude Opus 5.5 and GPT-6 Astra, which scored near zero due to safety refusals.
  • Cybench: The model solved 93% of the 40 exercises in this security competition dataset.

Safety and Robustness

  • B3 AI Security Benchmark: ML4 resists 93.3% of attacks, demonstrating high robustness to indirect prompt injections.
  • KORA Benchmark: ML4 achieved a score of 1.691 (where 2 is exemplary), the highest among the tested open-source models.
  • Malicious Requests: Despite its high cyber capabilities, ML4 maintains a higher average refusal rate for malicious requests (via JailbreakBench, StrongREJECT, and AgentHarm) than other open-source models.

Agentic Coding and Professional Workflows

ML4 demonstrates high proficiency in software engineering and complex business automation.

Coding and Agentic Benchmarks

  • Coding Agent Index: ML4 achieved a combined score of 49.8%, placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
  • Specific Benchmarks: It scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.
  • Human Evaluation: In a blind study by Surge AI, ML4 ranked second out of five models with a score of 3.74, trailing only Claude Opus 5 (4.22).

Business Automation

  • AutomationBench: ML4 scored 59.9% across 657 business workflows (including Gmail, Slack, and Salesforce), outperforming Kimi K3 and DeepSeek V4 Pro.
  • AA-Briefcase: For long-horizon knowledge work, ML4 reached 1,393 Elo, surpassing DeepSeek V4 Pro.

Multimodal and Scientific Capabilities

ML4 introduces a significant advancement in visual understanding and scientific reasoning.

Visual Grounding

ML4 is capable of reasoning across complex documents, charts, and gigapixel satellite imagery. In visual grounding tests, it surpassed GPT-6-Astra on the Dense 200 benchmark (42% vs 41%).

Science and Mathematics

  • SciCode-Verified: ML4 is state-of-the-art among open-weight models for implementing complex scientific workflows in code.
  • STEM Performance: Internal evaluations show ML4 is preferred over GLM-5.3 in CAD and STEM tasks, while remaining competitive in finance and coding.
  • Applied Math: The model can generate full Hartree–Fock simulations and handle frontier theoretical physics tasks.

Knowledge Work: Legal and Finance

ML4 is optimized for high-precision professional tasks. Third-party evaluations by vals.ai indicate that ML4 exceeds GPT-6-Astra in representative legal and financial tasks. Additionally, ML4 outperforms all open-source models on HarveyAI’s Legal Agent benchmark.

Reinforcement Learning at Scale

Mistral utilizes a composable Reinforcement Learning (RL) library to adapt the model as it evolves. This system allows a single training run to combine diverse tasks—from safety alignment to long-horizon tool use—using shared scaffolds like code sandboxes and web search.

At a scale of 3,000 GPUs, a single training run produces approximately 33 billion tokens per day, with 16 billion being trainable completion tokens. Mistral reports that the RL process is currently "in flight" and has not yet saturated, suggesting further capability gains are expected as compute scales.

Community Perspective

Community reactions from Hacker News highlight a mix of excitement and critical observation:

"Open weight, European, competes with GLM-5.3 on cybersecurity. What's not to like?"

Some users expressed a desire for more non-Chinese open models to challenge the current dominance of Chinese open-weight models, while others noted the efficiency of smaller models like DeepSeek v4.1 Flash in comparison to the massive scale of ML4.

Sources

Related

  • Dispatch
    Mistral Large 4 Release Notes

    Mistral AI has launched a public preview of Mistral Large 4, a 1 trillion-parameter natively multimodal open-weight model designed for frontier performance in cybersecurity, agentic coding, and professional knowledge work.

  • Dispatch
    Mistral Large 4 Release Notes

    Mistral AI has released Mistral Large 4, an open-weight multimodal model featuring a 1.05T parameter Mixture-of-Experts architecture optimized for cybersecurity and visual grounding.

  • Dispatch
    Mistral Large 2 release notes / what's new

    Mistral AI has released Mistral Large 2, a 123B parameter model designed for high throughput single-node inference with competitive performance in coding, reasoning, and multilingual capabilities.

  • Dispatch
    Pixtral Large Release Notes

    Mistral AI has released Pixtral Large, a 124B open-weights multimodal model that achieves state-of-the-art performance on MathVista, DocVQA, and VQAv2 while maintaining the text capabilities of Mistral Large 2.

  • Dispatch
    Mistral 3 Release Notes: Mistral Large 3 and Ministral 3 Family

    Mistral AI has released Mistral 3, featuring the frontier-scale Mistral Large 3 MoE model and the edge-optimized Ministral 3 family (3B, 8B, 14B), all under the Apache 2.0 license.