Mistral announced on October 6 the preview of Mistral Large 4 (ML4), nicknamed internally "Le Chonk" — a 1.05 trillion-parameter model with 52 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in European data centers. On October 29, the model weights were released for open download, marking the most ambitious moment in Europe's AI sovereignty strategy.
What makes ML4 different is not just scale, but strategic positioning. Mistral is not competing with OpenAI and Anthropic on the same terrain: it is building an open alternative for governments, cybersecurity firms, and organizations that need total control over their AI models. The decision to release weights for download, rather than offering API-only access, is deliberate — defense companies, financial institutions, and government agencies can run the model on their own infrastructure, eliminating dependence on external providers.
The model's architecture follows the Mixture of Experts (MoE) approach that has become standard in the industry, but with an important nuance: of the 1.05 trillion total parameters, only about 52 billion are activated per inference. This means the model does not mobilize all its capacity for each request, resulting in significant efficiency. Mistral itself states that ML4 uses 47% fewer tokens per task, on average, than Chinese model GLM-5.3.
Cybersecurity is ML4's primary focus. On the Artificial Analysis Cyber Index, which evaluates models' ability to find and fix security vulnerabilities in real software, ML4 ranks among the five strongest models globally and leads among open-weight models developed outside China. In one test, the model was asked to reproduce a real vulnerability in open-source software and then patch it — ML4 scored 82%, the highest of any model. In Cybench, a set of 40 exercises drawn from security competitions, it solved 93% of challenges.
But cybersecurity is only part of the picture. ML4 is also multimodal (accepts text and images as input, produces text output), has a context window of up to 1 million tokens, and was trained in over 160 languages, including all official languages of the European Union. On agentic coding benchmarks, it achieved 59.9% on AutomationBench — 657 business workflows across apps like Gmail, Google Sheets, Slack, and Salesforce — surpassing Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro.
The launch strategy was equally deliberate. Mistral released a preview on October 6, spent three weeks testing the model with cybersecurity partners, government authorities, and customers under real-world conditions. During that period, the weights were refined with additional reinforcement learning and customer feedback. The result was the introduction of two additional reasoning modes (low, medium, and high), allowing users to adjust the depth of reasoning according to the task. On Terminal-Bench 4.0, which tests AI agents' ability to perform coding tasks in a terminal, ML4's score rose from 28% to 41% after this refinement.
The strategic question behind it all is European technological sovereignty. Mistral's valuation is approximately US$ 24 billion — significantly below competitors like OpenAI (US$ 800 billion) and Anthropic (US$ 45 billion). However, the company argues that despite the resource disparity, ML4 is currently the best open model built in the West. The model is designed for scenarios where trusting closed providers is problematic: proprietary software source code, satellite imagery analysis for forest fire detection, power line risk monitoring, and even converting 2D CAD blueprints into usable 3D models.
ML4's release comes at a moment of increasing geopolitical tension in AI. With American restrictions on advanced chip exports to China and the race to develop independent alternatives, Europe sees open-weight models as a strategic tool. The ability to run ML4 locally means European governments do not need to trust their critical security systems to foreign providers' infrastructure — a distinction that has become increasingly relevant in a world where sensitive data is both valuable and dangerous.
What remains to be seen is whether ML4's efficiency in compute resources will be sufficient to sustain its market position in the coming months. The race for frontier models is accelerating — Google launched Gemini 4 Carbon a week before Mistral's weights were released, and each launch sets a new baseline. ML4 is not the end of the story, but the beginning of a new generation of specialized Mistral models designed for specific industries. If the company can maintain its pace of innovation, "Le Chonk" may be just the first chapter of an ambitious European attempt to redefine who controls the world's most advanced artificial intelligence.
Sources: Mistral, Euronews, CNBC
✓ Independent sources cross-checked and verified before publishing
