Mistral releases weights of trillion-parameter model ML4 for self-hosted deployment in cyber defense

Open weights instead of a cloud service

France's Mistral has published the weights of its flagship model Mistral Large 4 (ML4), known by the nickname 'Le Chonk'. Organizations can now download the model and run it on their own infrastructure rather than relying solely on a hosted service from an external provider; the weights are available on Hugging Face and via Mistral's API. According to TechCrunch, the model was first unveiled earlier this month, and euronews specifies that the ML4 preview was released on October 6, and over the past three weeks the company has continued reinforcement learning, testing the model with customers and partners in real-world conditions.

Claimed performance

Mistral claims that ML4 is 'thebest open agentic coding tool in the world' and comparable in performance to the Chinese open model GLM-5.3, previously considered one of the best in the world, especially in cybersecurity tasks. According to the company, ML4 uses on average 47% fewer tokens per task, and its score on the Terminal-Bench 4.0 test rose from 28% to 41% after additional training. The model contains 1.05 trillion parameters, of which about 52 billion are activated, and offers three reasoning modes — low, medium and high. Among the claimed applications: converting 2D drawings into 3D models via CAD, analyzing large volumes of data for financial decisions, and detecting forest fires and crop conditions from satellite imagery. These figures come from Mistral itself, and there is no independent verification in the publications.

Cyber defense and sovereignty

The model's positioning is tied to growing cyber threats in Europe — euronews mentions the recent PHP vulnerabilitydiscovered with the help of AI, as well as cyberattacks on Belgium's finance-tax agency and the Norwegian government. Mistral's vice president of science, Pierre Stock, told TechCrunch that open weights mean access to state-of-the-art AI for defense 'that by definition cannot be cut off', and that the model can be deployed locally to detect and respond to incidents in real time. The context is a June episode when the Trump administration ordered Anthropic to suspend access to the Fable 5 and Mythos 5 models for foreign nationals; restrictions on Fable 5 were later lifted, while access to Mythos 5 remained restricted. Stock also acknowledged the risks associated with incidents involving 'rogue AI agents', but warned against overreacting: 'We don't want to scare people. We believe this is a solvable problem.'

Resources and dependence on Nvidia

Mistral CEO Arthur Mensch said that 'open source is vital so that state-of-the-art AI does not end up concentrated in the hands of a few players', and that over three years the company has raised about $6 billion, whereas other labs have raised 'hundreds of billions'. According to euronews, Mistral's estimated valuation is about $24 billion (€20.9 billion), which is noticeably lower than the valuations of OpenAI and Anthropic. The model was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers: European infrastructure gives customers more operational control, but the dependence on American chips, as euronews notes, illustrates Europe's continuing reliance on foreign suppliers at key links in the AI chain. Mistral added that ML4 is part of a new generation of 'all-modality' models and that it will present more models from this generation in the coming months.