Mistral Releases Open Weights for Mistral Large 4: 47% Fewer Tokens per Task Than GLM-5.3
Mistral has released the weights for Mistral Large 4 (ML4), an open trillion-parameter model nicknamed Le Chonk. Anyone can download it and run it on their own servers. According to the company, ML4 performs no worse than China's GLM-5.3 in cybersecurity but uses 47% fewer tokens per task, which makes it cheaper to run. Security teams should therefore test ML4 on their own logs and code and compare the cost per task.
AI-processed from TechCrunch; edited by Hamidun News
France's Mistral has published the weights of its largest model, Mistral Large 4, or ML4, for download. It is an open model with a trillion parameters (the internal "settings" where a neural network stores what it knows), and inside the company it is called Le Chonk. Mistral compared it with China's open model GLM-5.
3, which was previously considered one of the best open models in the world, especially for cybersecurity. According to Mistral, the quality is comparable, but ML4 uses 47% fewer tokens per task. Tokens are the "units of text" you pay for when working through an API, and they load the hardware when you run the model yourself.
In other words, the same work comes out noticeably cheaper. For those choosing AI to protect infrastructure, this is a reason to compare what is more cost-effective: a paid closed API or your own model on your own servers.
What "open weights" means
The weights are a file of numbers in which everything the model can do is "recorded" after training. If the file sits on your disk, the neural network runs without contacting anyone else's cloud: data does not leave your environment, and the provider cannot cut off your access. The ML4 weights are published on Hugging Face, and the model is also available through Mistral's own API.
Mistral's Vice President of Science, Pierre Stock, explained this using a defense example: a company or government gets advanced AI to fight cyberattacks, and that access cannot be cut off because the model has open weights. It can be deployed on-prem, that is, inside one's own infrastructure. According to Stock, ML4 can be used to detect major security incidents, coordinate responses, and react to them in real time.
CEO Arthur Mensch repeats the company's long-standing position: AI should be a "tool for expanding capabilities, not for domination." Mistral claims it reaches the level of the best models with a "fraction of the compute and capital" that closed labs such as OpenAI and Anthropic spend.
What about the numbers
The main argument is savings. Imagine an agent that parses logs and checks suspicious code and configurations: it takes dozens of steps in a row. If each task uses 47% fewer tokens, the bill and response time fall with the same result. For comparison, Mistral chose GLM-5.3 specifically, which until this release was considered the strongest open model for cyber tasks.
The second argument is ranking. Mistral states that ML4 outperforms all other open models in agentic programming on the Coding Agent Index from Artificial Analysis. For security this matters, because finding vulnerabilities is closely tied to a model's ability to read and write code.
Limitations and conclusion
First, all the figures so far come from Mistral itself, and this article includes no independent verification. Second, open weights do not mean cheap deployment: a model with a trillion parameters requires serious hardware, and the token savings have to outweigh the costs of servers and support. A small team may find an API simpler, while a bank, telecom operator, or government agency with strict data requirements often needs its own deployment simply because of the rules.
Third, openness has a flip side, and Stock acknowledges it: attackers can use the same capabilities. He argues that vulnerability-finding tools should be made available to everyone as quickly as possible to strengthen the ecosystem. He called the alarm around recent incidents involving "rogue" AI agents exaggerated: the problem is complex but solvable.
The practical takeaway is simple. If you are responsible for security or build products for regulated industries, run ML4 as a short pilot. Test it on your own logs, code, and typical incidents, and compare the cost per task, analysis quality, and speed against your current model. The claimed 47% token savings are easy to verify on your own data, and the results will show whether switching is worth it.
- Nov 21, 2026Microsoft Lets Businesses Fine-Tune MAI Models on Their Own Data in Azure
- Nov 21, 2026Mistral Opens Singapore Hub for Asia, Plans to Grow to 200 Staff by 2027
- Nov 21, 2026Magnific Hits $300M in Annual Revenue on Other Companies' AI Models, Without Investors
- Nov 21, 2026Paris-Based Genesis AI Raises $500M at $3.5B Valuation for Eno Robots
Key Questions Answered
What "open weights" means?
The weights are a file of numbers in which everything the model can do is "recorded" after training. If the file sits on your disk, the neural network runs without contacting anyone else's cloud: data does not leave your environment, and the provider cannot cut off your access. The ML4 weights are published on Hugging Face, and the model is also available through Mistral's own API.
What about the numbers?
The main argument is savings. Imagine an agent that parses logs and checks suspicious code and configurations: it takes dozens of steps in a row. If each task uses 47% fewer tokens, the bill and response time fall with the same result. For comparison, Mistral chose GLM-5.3 specifically, which until this release was considered the strongest open model for cyber tasks. The second argument is ranking.
What are the key details on Limitations and conclusion?
First, all the figures so far come from Mistral itself, and this article includes no independent verification. Second, open weights do not mean cheap deployment: a model with a trillion parameters requires serious hardware, and the token savings have to outweigh the costs of servers and support. A small team may find an API simpler, while a bank, telecom operator, or government agency with strict data requirements often needs its own deployment simply because of the rules.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.