French artificial intelligence startup Mistral on Tuesday introduced a preview of its largest open-weight AI model yet, positioning it as a high-performance alternative to competing Chinese models.
According to AFP and Reuters, Mistral announced Mistral Large 4, a new open-weight model whose trained weights can be downloaded, deployed on private servers, and fine-tuned for specific use cases. Nicknamed “Le Chonk,” the model is scheduled for public release on Oct. 27.
The launch is Mistral's first major product announcement since May. It also follows the company's disclosure last month that it had raised EUR 3 billion, or about KRW4.7 trillion, in a Samsung-led funding round—the largest such investment in a European technology company to date.
Mistral Large 4 is a multimodal mixture-of-experts model with more than 1 trillion total parameters, while activating 49 billion parameters during inference. The company said it can process up to 1 million tokens of context at once and delivers leading performance among open-weight models in cybersecurity, finance, legal work, and industrial applications.
Mistral is targeting governments and enterprises that place a premium on data security and control over their AI systems through its open-weight strategy.
“From financial services to manufacturing and the public sector, businesses and institutions can directly manage their intelligence, data, compute, and operations on their own critical infrastructure—without exposing their most valuable knowledge externally,” Mistral said in a statement.
Leading U.S. AI companies including OpenAI, Google, and Anthropic generally follow closed-model strategies, retaining strict control over how their models are accessed and used. Chinese AI companies such as DeepSeek and Moonshot, by contrast, have largely embraced more open approaches.
At a launch event in Abu Dhabi, Mistral CEO Arthur Mensch said the newly announced model outperforms Chinese rivals in certain areas, including cybersecurity.
“The notion that Europe cannot compete is simply not true,” Mensch said.
Pierre Stock, Mistral's vice president of science, told Reuters that the new release was the most powerful open-weight model available to date. He added that the model had attempted to move beyond its test environment, but Mistral had anticipated and blocked the behavior.
Mistral said the model was trained on more than 160 languages and can both accept prompts and generate responses in each of them.
