Mistral Large 4
Introducing Mistral Large 4: "Le Chonk"
Mistral is proud to announce the public preview of Mistral Large 4 (ML4). This model represents a massive leap in our pursuit of frontier AI, combining immense scale with the flexibility of open weights.
🚀 Technical Overview & Performance
ML4 is a natively multimodal powerhouse designed for the most demanding enterprise tasks. Its architecture is defined by a massive scale:
The model is engineered for exceptional proficiency in:
- Advanced Coding and software engineering.
- Agentic Workflows (autonomous task execution).
- Multimodal Understanding (integrating text and vision).
While it competes with the world's most powerful open-source models, ML4 merely matches significantly outperforms any open-weight alternative originating from the US or Europe. In specialized fields like visual grounding, it even eclipses the capabilities of leading closed-source frontier models.
🇪🇺 Forged in Europe: Sovereign AI
ML4 isn't just a model; it's a statement on digital autonomy. It was trained from the ground up using 3,800 NVIDIA Grace Blackwell GPUs housed within Mistral's own European datacenters.
"Our goal is to provide state-of-the-art performance in critical verticals through open weights, ensuring customers maintain absolute control over their AI infrastructure."
Why Sovereignty Matters
In high-stakes sectors like cybersecurity, relying on a closed provider is a risk. Provider-level refusals can hinder block essential vulnerability research or incident response. If a capability is revoked mid-crisis, it becomes a security liability. ML4 solves this by offering:
- Open Weights: Full autonomy over deployment.
- Self-Deployment: Run on-premise or via private cloud.
- Legal Compliance: A European deployment operated end-to-end under EU law.
Global Reach, Local Nuance
ML4 is truly global. A vast portion of its training data was multilingual, covering over 160 languages, including every official language of the European Union.
🛡️ Deep Dive: Cybersecurity Capabilities
ML4 is positioned as one of the most potent AI tools for security professionals. It excels where closed models often fail due to overly restrictive safety filters.
Benchmarks & Validation
| Benchmark | ML4 Performance | Note |
|---|---|---|
| AA Cyber Index | Top 5 Globally | Leads non-Chinese open-weight models |
| Vulnerability Patching | 82% | Highest score for reproducing & fixing flaws |
| Cybench | 93% | Solves nearly all 40 security competition tasks |
The "Refusal" Gap:
While ML4 thrives, closed models like Claude Opus 5.5 and GPT-6 Astra often score near zero on these tests because they refuse to engage with "potentially harmful" security tasks. However, defending a system requires the ability to prove a flaw exists—a task ML4 handles with ease.
Beyond Training: Internal tests show ML4 is highly effective at:
- Analyzing complex malware.
- Prioritizing critical vulnerabilities.
- Generating precise detection rules.
💻 Agentic Coding & Engineering
ML4 is a master of the terminal and the repository. It doesn't just write snippets; it understands entire codebases.
Coding Metrics
- DeepSWE v1.1:
61.7% - SWE-Atlas-QnA:
59.4% - Terminal-Bench 4:
28.3%
With a Coding Agent Index score of 49.8%, ML4 surpasses competitors such as DeepSeek V4 Pro 0813 and Qwen3.8 Max.
# Example of ML4's agentic capability:
# Analyzing a repo, finding a bug, and applying a patch
$ ml4-agent --analyze ./src/auth_module
$ ml4-agent --fix "CVE-2026-XXXX" --apply
# Result: Patch applied successfully. Tests passing.
🛠️ Development & Roadmap
ML4 was developed in collaboration with global leaders in pharmaceuticals, finance, logistics, shipping, and the public sector. It utilizes the same RL (Reinforcement Learning) and customization environment found in Mistral Forge.
The Path Forward
Current Status & Next Steps:
- Launch Public Preview API
- Complete red-teaming with cybersecurity leaders
- Release detailed architecture documentation
- Publish expanded benchmarks
- Release open weights to the public
We invite you to explore the preview API and share your experiences. ML4 is the foundation for a new era of specialized Mistral models.