Mistral Large 4 Release Notes

Mistral AI has announced the public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters (49 billion active parameters). ML4 is designed to push the frontier of open-weight performance, specifically targeting high-stakes enterprise workloads in cybersecurity, finance, law, and agentic software engineering.

Technical Specifications and Infrastructure

Mistral Large 4 was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs within Mistral's own European datacenters. The model's training data was heavily multilingual, covering over 160 languages, including all official languages of the European Union.

Key architectural and deployment details include:

  • Parameter Count: 1 trillion total parameters, with 49 billion active parameters.
  • Deployment: Available via preview API on Mistral Studio, with weights scheduled for release by the end of the month.
  • Sovereignty: The model is operated end-to-end by Mistral in Europe, independent of other digital service providers and governed by European law.

Cybersecurity Capabilities

ML4 is positioned as a state-of-the-art open-weight model for cybersecurity, prioritizing the ability to perform vulnerability research and incident response without the restrictive refusals common in closed-source models.

  • Performance Benchmarks: ML4 ranks among the top five models globally on the Artificial Analysis Cyber Index. It achieved an 82% score on a test requiring the reproduction and patching of real vulnerabilities in open-source software—the highest score of any model. Additionally, it solved 93% of challenges in Cybench.
  • Competitive Advantage: Mistral notes that closed models like GPT-6 Astra and Claude Opus 5.5 often score near zero on these specific tests due to safety filter refusals. ML4 provides the capability to analyze malware, prioritize vulnerabilities, and write detection rules while remaining deployable on private clouds or on-premise for auditability.

Agentic Coding and Software Engineering

ML4 demonstrates high proficiency in repository understanding, terminal workflows, and autonomous software engineering tasks.

  • Coding Benchmarks: The model scored 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.
  • Human Evaluation: In a blind evaluation by Surge AI, ML4 Preview ranked second out of five models with a score of 3.74, trailing only Claude Opus 5 (4.22) and outperforming Kimi K3, GLM-5.3, and GLM-5.2.
  • General Automation: On AutomationBench (657 business workflows), ML4 scored 59.9%, surpassing Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro.

Multimodal Understanding and Science

ML4 introduces a significant advancement in visual reasoning, enabling the model to process complex documents, charts, and geospatial imagery.

  • Visual Grounding: ML4 surpassed GPT-6-Astra on the Dense 200 benchmark (42% vs 41%). Its capabilities include inspecting gigapixel satellite imagery for disaster response and verifying mechanical parts in technical drawings.
  • Scientific Workflows: The model is state-of-the-art among open-weight models on SciCode-Verified. It can generate full Hartree–Fock simulations for chemistry and perform complex applied-mathematics tasks relevant to theoretical physics.

Professional Knowledge Work

ML4 is optimized for high-precision tasks in legal and financial sectors.

  • Legal and Finance: Third-party evaluations by vals.ai indicate that ML4 exceeds GPT-6-Astra in representative legal and financial tasks. It also outperforms all open-source models on HarveyAI’s Legal Agent benchmark.
  • Financial Analysis: The model is capable of multi-step corporate finance challenges, including searching through public company filings via EDGAR and European databases to synthesize information.

Model Safety and Robustness

ML4 implements advanced safety measures while maintaining its utility for cybersecurity researchers.

  • Attack Resistance: On Lakera’s B3 AI Security Benchmark, ML4 resists 93.3% of attacks, the highest score among its competitors. It also shows high robustness to indirect prompt injections compared to GLM-5.2, GLM-5.3, Kimi-K2.6, Kimi-K3, and DS-V4-Pro-0813.
  • Responsible Engagement: ML4 achieved a KORA Benchmark score of 1.691. Despite its high cyber-capability, it maintains a higher average refusal rate for malicious requests (measured via JailbreakBench, StrongREJECT, and AgentHarm) than other open-weight models.

Reinforcement Learning (RL) at Scale

Mistral utilizes a composable RL library to adapt the model as it evolves. This system allows a single training run to combine diverse tasks such as safety alignment, factuality, and long-horizon tool use.

  • Training Scale: Using 3,000 GPUs, a single training run produces approximately 33 billion tokens per day, with 16 billion being trainable completion tokens.
  • Infrastructure: The pipeline is optimized for long trajectories and high rollout budgets (millions of tokens) to minimize off-policy drift and ensure stable RL over long horizons.

Sources