Mistral has publicly released the weights for Mistral Large 4, its trillion-parameter AI model aimed at cybersecurity, coding and enterprise workloads. The October 29 release follows an October 6 preview and adds three reasoning settings—low, medium and high—alongside improved results on an agentic coding benchmark.
The model, also known as ML4 and nicknamed Le Chonk, can be downloaded and deployed on an organisation’s own infrastructure. Its weights are available through Hugging Face under a modified MIT-style license, giving customers an alternative to accessing a model exclusively through a provider’s hosted service.
That distinction is central to Mistral’s security pitch. Running a model internally can let organisations process sensitive source code and other operational data without sending it to an external inference service. Holding the weights also reduces dependence on a provider’s continued availability, although customers remain responsible for their deployment, security controls and compliance with the license.
ML4 has roughly 1 trillion total parameters, with 52 billion active for each token it processes. That sparse design limits the computation used at each step, rather than activating the entire model. The reported minimum deployment configuration is four Nvidia B200 or B300 GPUs. The model also supports a 1 million-token context window and more than 160 languages.
The three reasoning modes give developers a way to adjust the model’s reasoning effort to the task. Such settings are intended to balance response speed and resource consumption against the demands of more complex work. They are separate from the context window, which determines how much material the model can consider in a request.
On Terminal-Bench 4.0, which evaluates agents carrying out tasks in a terminal environment, ML4’s score rose from 28% in preview to 41% at release. Mistral attributes the improvement to reinforcement learning informed by feedback from government and corporate users during the preview period.
The broader benchmark picture is more mixed than that single improvement suggests. On the Artificial Analysis Coding Agent Index, ML4 scored 48, matching Codex GPT-6.1 Terra and trailing GLM-5.3’s 49. Mistral says ML4 uses 47% fewer tokens per task on average than GLM-5.3 in that comparison. Token consumption is relevant to operating costs, but does not by itself establish the total cost of running either model.
Cybersecurity is another focus of the release. Mistral is positioning ML4 for finding vulnerabilities, developing proof-of-concept inputs and producing patches. Reported CyberGym-E2E results fall between 79% and 82%. The company also says the model has high refusal rates for malicious requests. Those benchmark results and safety claims do not establish how reliably it will perform across unfamiliar production systems or adversarial prompts.
Open weights are not synonymous with unrestricted use or a fully open development process: the accompanying license still matters, and access to model parameters does not necessarily include training data or the complete training pipeline. For organisations that do not want to operate the hardware themselves, hosted API access is priced at approximately $0.70 per million input tokens and $2.10 per million output tokens.
Sources: Technobezz, ZDNET, Crypto Briefing