What changes with the release of the weights
Mistral has published the weights of its open-source trillion-parameter model, Mistral Large 4, nicknamed Le Chonk and unveiled earlier this month. According to TechCrunch's account, the release allows companies and public administrations to download and install the system on their own. The weights, that is, the numerical values resulting from training that make the model work, are available on Hugging Face and via API.
The central point, for the company, is control: an open-weight model cannot be revoked or made inaccessible by the vendor. It is the argument Mistral sets against labs that distribute closed models such as OpenAI and Anthropic, in an increasingly heated comparison between the two approaches. The company argues that companies and governments should not depend on external parties to access advanced capabilities.
The bet on cybersecurity
Mistral presents ML4 as a sovereign cyber defense tool: identifying, analyzing and responding to threats on code and IT infrastructure. Pierre Stock, vice president of scientific research, told TechCrunch that opening the weights allows the model to be run on-premise, within the organization's own ecosystem, and to respond in real time to serious incidents without access being interrupted. CEO Arthur Mensch summed up the position with the idea that AI should be a tool of emancipation and not of domination.
The company claims performance comparable to GLM-5.3, a Chinese open-weight model considered among the best in the world, especially in security tasks, with 47 percent fewer tokens per task and therefore a lower cost. It also states that ML4 surpasses other open models in agentic coding tasks according to Artificial Analysis's Coding Agent Index. These are the company's claims, not independent verifications reported in the material.
Stock also touched on the issue of AI agents out of control: he acknowledged the risks while urging against alarmism and calling the problem difficult but solvable. The spread of defensive capabilities, he added, must be extended rapidly because malicious actors could use their own technology to attack.
What remains to be verified
The release is a confirmed fact. What remains to be measured independently are the claimed performance, the token savings and the comparison with GLM-5.3, as well as the actual cost-effectiveness compared with closed labs. The material does not indicate license terms, hardware requirements for on-premise execution or adoption timelines. The scale of the challenge to OpenAI and Anthropic, and its effects on the market, are therefore to be followed in the coming months.
