French artificial intelligence company Mistral AI launched Mistral Large 4 (ML4) on October 6, a multimodal model with one trillion parameters nicknamed 'le Chonk' — a humorous reference to internet culture's depiction of an extremely fat cat. The release positions the European company as a direct alternative to American closed models and Chinese open-weight models, aligning with what French President Emmanuel Macron has described as 'a third way in AI.'
The model, accessible via API during a public preview phase, activates 49 billion parameters per query by employing a hybrid sparse mixture-of-experts (MoE) architecture. This means that while the model contains one trillion parameters in total, only a fraction is activated during inference, making the model significantly more hardware-efficient than equivalent models that activate all their parameters on every request. The preview pricing is set at $1.36 per million input tokens and $4.18 per million output tokens, significantly cheaper than competitors like Claude Opus 5.5 and GPT-6 Astra.
ML4 was trained from scratch over approximately two months on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. According to the company, this represents two to three times less hardware than Chinese competitors and significantly less than closed-source American rivals. The company claims to have trained the model in more than 160 languages, including all official languages of the European Union.
Preliminary benchmark results show ML4 achieving 61.7% on DeepSWE v1.1, a long-horizon software engineering benchmark; 28.3% on Terminal-Bench 4; 59.4% on SWE-Atlas-QnA; and 59.9% on AutomationBench, which tests enterprise workflows across apps like Gmail, Google Sheets, Slack, and Salesforce. Its Combined Coding Agent Index score of 49.8% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In cybersecurity, the model scored 82% on a test that asks a model to reproduce a real vulnerability in open-source software and then patch it — the highest score of any model. It also solved 93% of the challenges in Cybench, a set of 40 exercises drawn from security competitions.
Mistral plans to release the model weights by the end of October, following a three-week testing period with developers, cybersecurity leaders, and state authorities. The weights will be licensed under a custom Mistral license. During the preview period, the company will continue reinforcement learning and refine the final checkpoint before public release.
The launch comes at a crucial moment for the European AI industry. In September, Mistral announced a €3 billion Series D funding round with a post-money valuation above €21 billion — the largest equity fundraising ever completed by a European technology company, according to Reuters. The company now claims to support more than 125 global enterprises, including Airbus, ASML, and HSBC.
Mistral's strategy combines open weights with a sovereign enterprise stack. The model was optimized specifically for cybersecurity, programming, and chip design — sectors fundamental to two of its main backers: Dutch giant ASML, which led its Series C, and Samsung, which led its Series D last month. The company is betting that model weights will become increasingly commoditized while the highest-value business shifts toward systems built around them.
Sources: Mistral AI, CNBC, WIRED
✓ Independent sources cross-checked and verified before publishing
