Mistral open-sources trillion-parameter ML4 model aimed at cybersecurity defense

Mistral publishes the weights of a trillion-parameter model

French AI firm Mistral has released the trained weights of Mistral Large 4 (ML4), a trillion-parameter model the company first showed earlier this month under the informal nickname "Le Chonk." According to TechCrunch, the weights are available for download on Hugging Face and accessible through Mistral's API, which means businesses and government agencies can now run the model on systems they control rather than depending on a hosted endpoint.

Performance and cost claims

Mistral claims ML4 matches the performance of GLM-5.3, a Chinese open-weight model previously considered among the strongest in the open category, particularly on security-related work. The company also says its model is cheaper to operate, consuming 47 percent fewer tokens per task than GLM-5.3. On the Coding Agent Index run by the benchmarking site Artificial Analysis, Mistral says ML4 ranks above every other open model for agentic coding tasks. These figures come from Mistral itself and were not independently verified in TechCrunch's reporting.

CEO Arthur Mensch has framed the release in ideological terms, saying that AI should "be a tool for empowerment, not dominance" and arguing that Mistral delivers leading performance while spending far less on compute and capital than closed laboratories.

Cyber defense and the sovereignty pitch

The company is steering the model squarely toward cybersecurity. In a statement reported by TechCrunch, Mistral said enterprises and public sector institutions can now own frontier AI outright and use it to detect, analyze and respond to threats across their code and IT infrastructure.

Pierre Stock, Mistral's vice president of science, told TechCrunch that open weights give organizations access to frontier-level defensive capability that, in his words, "can't be cut off by definition" because the model is downloaded rather than rented. He described use cases such as identifying, coordinating and responding to major security incidents in real time, and noted that open weights allow companies to deploy the model on-premises inside their own environments.

Stock also made an ecosystem argument: AI tools are already helping security professionals find software bugs faster than before, and that defensive capability should be distributed as widely and quickly as possible, because malicious actors may try to weaponize the same techniques for attacks.

On rogue AI agents

Asked by TechCrunch about recent incidents involving rogue AI agents, Stock acknowledged the risk but argued against alarmism. "We don't want to fearmonger people. We think it's a problem that is solvable. It's a hard problem to have good security," he said.

Why it matters

The release sharpens the split between open-weight developers such as Mistral and closed-model providers such as OpenAI and Anthropic. By publishing the weights of a trillion-parameter system, Mistral is arguing that organizations should not have to rent frontier capabilities from a third party — a case that carries particular weight in cybersecurity, where control over data and infrastructure is a hard requirement for many buyers. If the performance and token-efficiency claims hold up, ML4 also increases the pressure on closed labs, since customers can now compare a downloadable, self-hostable frontier model against subscription APIs. Two caveats temper the headline: the benchmark numbers are vendor-supplied, and serving a model at this scale demands significant hardware, so in practice owning frontier AI remains most realistic for well-resourced enterprises and governments rather than smaller teams.