Mistral opens trillion-parameter ML4 weights, targeting cyber defence

Mistral is turning open weights into a sales pitch against closed labs and Chinese open-weight rivals. On 29 October 2026 it published the weights of Mistral Large 4 (ML4), a model with one trillion parameters that it first introduced earlier that month. TechCrunch reported the release at 10:00 AM PDT (17:00 UTC) the same day.

The weights are on Hugging Face and available through Mistral's API, and organisations can deploy the model on-premises. Mistral is aiming ML4 at sovereign cyber defence, for enterprises and governments that want to deploy and control the model on their own infrastructure. It also highlights agentic coding as a use case.

The performance claims come from Mistral's own comparisons and have not been independently verified. Mistral says ML4 performs comparably to GLM-5.3, a Chinese open-weight model, while using 47% fewer tokens per task. It also says ML4 outranks all other open models on Artificial Analysis' Coding Agent Index, and that it reaches state-of-the-art performance at a fraction of the compute and capital of closed labs.

Pierre Stock, Mistral's vice president of science, framed the open-weight case in one line: "Access that can't be cut off by definition because the model is open-weighted". Stock said Mistral wants to make this cyber capability available to everyone as soon as possible.

Some details remain unconfirmed. TechCrunch calls ML4 open-weight but names no licence, so the terms of use are not clear from the reporting. Benchmark claims beyond Mistral's own comparisons are unverified, and no timeline for future releases has been confirmed.

The biggest winners are enterprises and governments that need AI inside their own walls. ML4 gives them a trillion-parameter model they can host themselves, without depending on an API that a provider can switch off. That is the buyer Mistral is targeting with the sovereign cyber pitch.

Closed labs face the sharpest pressure. Mistral is measuring ML4 against a Chinese open-weight rival and stressing token use per task, a metric that matters to anyone paying for agentic coding. Watch three things: the licence terms, whether the benchmark claims hold up outside Mistral's own tests, and whether the cyber pitch wins real deployments with governments and large enterprises.