Mistral releases Large 4 open weights, aims 1-trillion model at cyber defence

French AI company Mistral AI released the weights of its flagship model, Mistral Large 4, to the public on Thursday, October 29, 2026, so that companies can download it and run it on their own servers. The company, which has nicknamed the model "Le Chonk", is pitching it mainly at cybersecurity teams that want to find and fix weaknesses in their software without sending sensitive code to an outside provider. The model is also available through Mistral's own online service.

Most of the strongest AI systems today are "closed": a company runs them in its own data centres and customers send their questions and documents over the internet. An "open-weight" model is different. The weights are the billions of numbers the model learned during training; publishing them lets any organisation run the model on its own machines, inspect how it behaves and adapt it, without its data leaving the building.

For security work, that difference is central. To find flaws, a model must read a company's source code, network settings and incident logs, which are among its most sensitive files. Many banks, government bodies and defence contractors cannot legally or safely send that material to an outside AI service. A capable model they can host themselves removes that barrier. Mistral has framed this as a question of "sovereignty", meaning control over one's own data and tools.

The size figures need a word of explanation. "Parameters" are the adjustable values inside a model. Large 4 uses a design in which only a slice of the model, about 49 billion of its roughly 1 trillion parameters, is used for each word it processes. That makes it cheaper and faster to run than its total size suggests, which is why it can fit on a handful of high-end chips rather than a full data centre.

There is also a safety question. A model that is good at finding software flaws could in principle help attackers as well as defenders. Mistral says the model refuses malicious requests at high rates. Those are the company's claims, and because the weights are public, outside researchers will be able to test them independently.

For India, where companies and government departments are under growing pressure to keep personal and critical data within their own control, self-hosted models of this class widen the choice beyond closed services run abroad.

Short-term: Security teams and developers can begin testing the model on their own systems straight away. Cloud and AI gateway services have already started listing it.

Long-term: The release raises the bar for open models outside China and the United States, and adds to pressure on closed providers to offer private, on-premises versions. It may also sharpen the debate on how much powerful security capability should be published openly.

Who is affected: Corporate and government security teams; software developers; companies in regulated sectors such as banking and health; AI providers competing with Mistral; and policymakers weighing the benefits and risks of open-weight AI.

Questions and Answers