Ing. Calogero Bono
Mistral has released the weights of Mistral Large 4, nicknamed Le Chonk, a model with roughly one trillion parameters. It did so at the end of October, after a public preview earlier that month. You can download it from Hugging Face and also use it via API. It matters because it is an open model that the French company positions for code and cybersecurity, and because nobody can switch it off remotely.
A few terms before we get into the facts. The weights are the numerical parameters learned during training, the ones that make the model work. Open-weight means you can download them, not that all the code is free. MoE (Mixture-of-Experts) is an architecture in which, for each token, only a small group of internal "experts" does the work. A token is a piece of text processed by the model. A benchmark is a standard test for comparing models.
What is Mistral Large 4?
According to the official Hugging Face page, it is a multimodal, multilingual MoE model. It accepts text and images and generates text. It has about 1T total parameters, 49B active, and a 1M-token context window.
The license is the Mistral Ecosystem Open-Weight License, a derivative of the MIT license with limits for those who want to offer the model as a service. Punto Informatico reports that it supports over 160 languages and was trained on 3,800 NVIDIA Grace Blackwell chips. ZDNET, instead, cites Mistral's statement of 4,000 GPUs for two months. The two figures do not match, so we report both as attributed.
What numbers does Mistral claim?
On coding, the Hugging Face page lists 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. A Coding Agent Index of 49.8%. According to ZDNET, Terminal-Bench 4.0 rose to 41% from the preview's 28%, one point above GLM-5.3. On the overall index the model scores 48 against GLM-5.3's 49.
Mistral says it competes with GLM-5.3 while using 47% fewer tokens per task, as reported by TechCrunch. On the cyber side, Hugging Face cites 82% on a test that asks the model to reproduce a real vulnerability and then fix it, and 93% on Cybench (40 challenges). Punto Informatico writes that only DeepSeek V4.1 Flash scores higher on the dedicated benchmark, while ZDNET says Le Chonk sits just above DeepSeek V4.1 Flash on CyberGym. The sources are not aligned on this point.
Why do open weights matter for security?
Mistral's argument is control. Pierre Stock, VP of science, told WIRED that AI will be like electricity, and you do not want someone able to cut it off. Closed models can block requests to analyze malware or vulnerabilities, because they look like attacks. With the weights in-house, usage policy and deployment stay with you.
There is a flip side. Rafe Pilling of Sophos calls the model dual-use technology, because it can be downloaded and adapted by attackers too. According to Mistral, the model refuses harmful requests 97% of the time (figure reported by Punto Informatico). Hugging Face states that its average refusal rate on cyber prompts is higher than any other open-weight model.
To understand the incident the press has mentioned, read our article on the Hugging Face case.
How do you use it in practice?
The model has day-0 support for transformers, vLLM and llama.cpp, with GGUF quantizations. There are also Inference Endpoints and Inference Providers. Developers can also reach it through Mistral Studio, API and OpenRouter.
The reasoning level can be low, medium or high. The default is high and, according to the page, reasoning cannot be turned off. The NVFP4 version for vLLM requires NVIDIA Blackwell GPUs, and the official example uses 8 GPUs. This is not a small server model. If you need AI that is more within reach, see the AI guide for SMEs.
Limits to know
Almost all the numbers come from Mistral. ZDNET reminds readers that benchmarks should be taken with caution. Sources diverge on who leads in cyber and on training resources. Hugging Face's ranking places the model fifth among open-weight models on the Artificial Analysis Intelligence index, while Mistral calls it the best in the world for open coding and cyber, according to WIRED. These are different claims.
No independent verification confirms that the model resists malicious use after fine-tuning. Real running costs and tests on Italian cases are missing. Moreover, the license restricts offering it as a service.
What changes for those who work with digital tech
For developers, there is now an open European alternative for coding assistants and security analysis, which can run in-house. For those handling sensitive data, on-prem deployment avoids sending code and logs to third parties. As Stock told TechCrunch, access cannot be revoked.
For a small company, though, you need serious hardware or a provider. A lot depends on the vendor. The regulatory framework matters too, so we refer you to the guide to the AI Act for developers. If you manage servers, managed hosting and servers with security updates remain the foundation, whatever model you use.
Our take
The move is interesting, and publishing the weights makes sense for defense. Someone investigating a vulnerability should not have the door closed on them in the middle of an incident. The geopolitical race tone convinces us less. Pilling's words are a good antidote, because risk depends on monitoring and controls, not on the model's flag.
Moreover, a trillion-parameter model is out of reach for almost no Italian SME. The concrete benefit, for now, goes to agencies, large companies and providers. We will wait for independent tests before calling it the best.
Frequently asked questions
Is Mistral Large 4 truly open source?
It is open-weight, not open source in the strict sense. The weights can be downloaded from Hugging Face. The license derives from the MIT but limits offering the model as a service.
Can it be used for free?
The weights are freely available, according to Punto Informatico. Running the model, however, costs hardware. The official example uses 8 GPUs.
Is it better than Chinese models?
It depends on the test. Mistral says it matches or beats models such as DeepSeek and Z.ai. ZDNET reports it one point behind GLM-5.3 on the overall index.
Can it be used to attack?
Yes, like any downloadable model. Mistral claims it refuses dangerous requests 97% of the time. After fine-tuning, however, its behavior could change.
Which languages does it support?
Over 160, according to Punto Informatico, including all European ones. It accepts text and images and generates text.
What to watch in the coming months, between independent tests and real costs
- Follow independent tests, such as Artificial Analysis, and do not rely only on the company's numbers.
- Calculate the running cost, in hardware or via API, before migrating a workflow. To start small, check out five AI uses for small businesses.
- Prepare defenses against abuse, starting with jailbreaks and prompt injection.
