Mistral Releases Open Weights for Large 4, Its Trillion-Parameter Model for Coding and Cyber Defense
French AI company Mistral has released the weights of Mistral Large 4, its largest model to date, allowing businesses, governments and researchers to download and run the system on their own infrastructure. TechCrunch reported the release on October 29, 2026, following a public preview of the model earlier in October. The move turns a hosted preview into a downloadable model, and it is aimed at organisations that want frontier AI capability without depending on another company’s servers.
Mistral nicknames the model “Le Chonk”. It describes Large 4, also shortened to ML4, as a natively multimodal model, meaning it can work with text and images. Mistral says the model has about 1 trillion parameters in total, with 52 billion active parameters, and that it was trained from scratch on 3,800 NVIDIA Grace Blackwell graphics processors in its own data centres in Europe. Mistral also says training data spanned more than 160 languages, including every official language of the European Union.
From preview to open weights
Mistral first launched Large 4 as a public preview on October 6, 2026, with a promise that the weights would follow by the end of the month. The weights are the trained numerical values inside a model. Publishing them is what makes self-hosting possible: an organisation can download the model, run it on hardware it controls, fine-tune it on its own data, and keep it running even if a hosted service changes or becomes unavailable.
TechCrunch reports that the weights are being made available on Hugging Face and through Mistral’s application programming interface. That combination matters for adoption. Teams with the infrastructure to self-host can download the model directly, while teams that prefer a managed route can still call it through an API.
What the model is built for
Mistral positions Large 4 for agentic coding and cybersecurity work. In cybersecurity, Mistral says the model ranks among the top five models globally on the Artificial Analysis Cyber Index, an independent evaluation of how well models find and fix security flaws in real software. On one test in that index, which asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral says Large 4 scores 82 percent, the highest of any model. Mistral also says the model solves 93 percent of the challenges in Cybench, a set of 40 exercises drawn from security competitions.
That positioning reflects a practical problem Mistral highlights in its announcement. Defending software often starts with proving that a flaw is real, and safety filters in some closed models can refuse that kind of work. Mistral argues that open weights and self-deployment let organisations run advanced security work under their own policies, including work on private cloud or on-premise systems.
On coding, TechCrunch reports that Mistral says Large 4 outranks all other open models on Artificial Analysis’ Coding Agent Index. TechCrunch also reports that Mistral says the model is comparable in performance with the Chinese open-weight model GLM-5.3, while using 47 percent fewer tokens per task. Tokens are the units of text a model processes, so fewer tokens per task can mean lower cost for the same work.
Licence and pricing
The release uses a modified MIT licence, which Mistral calls the Mistral Ecosystem Open-Weight License. Under the terms reported for the release, the model is free to use, modify, fine-tune and commercialise, with one important condition: businesses that offer the model as a hosted model-as-a-service product must obtain a separate commercial licence from Mistral first. Any team planning to resell hosted access should read the licence terms before building on it.
For teams that prefer not to host the model themselves, Mistral’s API pricing is reported at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens.
Why it matters
The release sharpens a contrast in how frontier AI is distributed. Some leading models remain available only through a provider’s hosted service, which gives the provider control over access, pricing and safety filters. An open-weight release moves that control to the user. Mistral’s vice president of science, Pierre Stock, put the point directly to TechCrunch: “Access that can’t be cut off by definition because the model is open-weighted.” For governments and regulated companies, that continuity can matter as much as raw capability, especially during a security incident when losing access mid-response could itself become a risk.
There is also a flip side. The same capability that helps defenders reproduce and patch vulnerabilities could, in other hands, help attackers. Stock told TechCrunch that malicious actors might try to leverage their own technology to attack, and that defenders therefore need the same capability as soon as possible. That does not settle the debate about open-weight frontier models, but it explains why Mistral chose to release the weights rather than keep Large 4 behind a hosted gate.
For users of AI tools, the practical takeaway is narrower. Large 4 is not a consumer chatbot launch. It is a foundation model for developers, security teams and enterprises building coding agents, security tooling and internal AI systems. Its value will show up in the tools built on it and in security workflows that can now run on infrastructure their owners control.
