Mistral releases the weights of Mistral Large 4, a 1-trillion-parameter open model

What was released

Mistral published the model on Hugging Face as mistralai/Mistral-Large-4-1T-A52B. The page calls it a "1-trillion-parameter" model. Parametersthe numbers a model learns during training; more parameters usually means more capacity, and more memory needed to run it measure the model's size. The "A52B" in the repository name points to about 52 billion parameters being active for each piece of text, which is how a mixture-of-expertsa design where the model has many specialized sub-networks and uses only a few of them for each step, so it runs faster than its total size suggests model works. Mistral's announcement does not state the active count in its text, so treat 52B as the repository's label.

These are open weightsthe trained model files are published so anyone can download and run them, even though the training data and code may stay private. You can run the model with vLLM or SGLang, or use it through Mistral Studio, the Mistral API, OpenRouter, NVIDIA endpoints and Microsoft Foundry. API pricing is $0.70 per million input tokens and $2.10 per million output tokens.

The license, in short

Mistral calls it the "Mistral Ecosystem Open-Weight license", a modified MIT license. Most uses are allowed. The main exception: a business that offers the model as a hosted service to others needs a separate commercial license. Read the full text on the Hugging Face page before building a product on it.

Benchmarks reported by Mistral

Mistral reports 48% on Artificial Analysis's Coding Agent Index, which it describes as effectively tied with GLM-5.3 while using 47% fewer tokens per task. It also reports 41% on Terminal-Bench 4.0 and 84.5% on BrowseComp, and says it leads the open-weight models in Cognition's private FrontierCode 2.0 test with 19.7%.

These figures come from Mistral's announcement. FrontierCode 2.0 is private, so outsiders cannot reproduce it, and the preview score on Terminal-Bench was self-reported. Independent results usually arrive in the weeks after a release.

The security angle

Mistral pitches the model to defenders. It reports 51% on Artificial Analysis's Cyber Index, which tests finding, reproducing and fixing vulnerabilities, a 96.6% average refusal rate on malicious cyber requests, and 97.1% resistance on the B3 agent security benchmark. A separate cybersecurity variant is in preview for vetted partners, limited to systems they own or are authorized to test.

The same post says that in monitoring, the model refused black-box attacks on live targets, but that Mistral sees "a potential for uplift to conduct real attacks", and that it is changing its guardrails. With the weights public, anyone who fine-tunes the model can also train those guardrails away.

For a sense of what defenders deal with each week, our CVE Watch analysis of the exploited WordPress flaw shows a real attack window, and the Firefox 159 update shows how many memory bugs a single browser release can close.