Mistral Large 4 is now open: anyone can download the model

What's new

French AI developer Mistral has made its large model, Mistral Large 4, open. That means anyone can download the model's "weights," the numbers that make up the model's brain, and run it on their own machines. The model first arrived in a public preview on October 6, 2026, via Mistral Studio and Mistral's API. The announcement does not state the exact date the weights were released.

Large 4 has 1 trillion parameters in total, but uses only 52 billion at a time. It is a so-called mixture-of-experts model, in which only part of the model works on each task. That saves resources. It can take images as input, supports more than 160 languages (all of the EU's official languages are included) and has three levels of reasoning: low, medium and high. According to Mistral, it was trained from scratch in European data centers on 3,800 NVIDIA Grace Blackwell GPUs.

What's smart

The smart part is the freedom. The model can be self-hosted, so a company can keep its data within its own walls. It is also available through Mistral Studio, OpenRouter, Hugging Face, Microsoft Foundry and NVIDIA, and can be run with vLLM and SGLang or as NVIDIA NIM containers.

The license is a modified MIT license, the Mistral Ecosystem Open-Weight License. It permits use, modification, fine-tuning and commercial exploitation, but you may not offer the model onward as Model-as-a-Service, that is, as a paid service where others use the model itself.

Mistral also emphasizes cybersecurity. According to the company, the model refuses 96.6 percent of malicious cyber requests and scores 97.1 percent on the B3 Agent Security Benchmark, up from 93.3 percent in preview. For professionals, Mistral offers preview access to a cyber variant for penetration testing and red teaming, but only after screening and contractual safeguards. Mistral CEO Arthur Mensch is quoted as saying: "AI is software. It can be controlled." The outlet Runtimewire points out, however, that open weights make it easier to remove safeguards afterward.

Cheaper or better?

In Mistral's API, the model costs $0.70 per million input tokens (pieces of text), $0.07 for cached input tokens and $2.10 per million output tokens. A monthly price for ordinary users has not been disclosed.

On coding, Mistral reports 48 percent on the Artificial Analysis Coding Agent Index, where the model shares first place with GLM-5.3 and uses 47 percent fewer tokens. On Terminal-Bench 4.0 it gets 41 percent, 13 percentage points more than the preview, and 19.7 percent on FrontierCode 2.0. Surge Code evaluators preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. In the preview, it ranked second behind Claude Opus 5 in a human coding evaluation.

On the Artificial Analysis Cyber Index it scores 51 percent. At preview it placed among the top five and led among open models outside China. All figures are Mistral's own, and independent verification does not appear in the sources. Figures for the preview and the release also differ in several places.

What it's good at

Coding and tasks where the model works through multiple steps on its own are among its strongest points, according to Mistral's figures. The company also cites 73.3 percent on FinWorkBench and 84.5 percent on BrowseComp, which measure financial work and web search.

Cybersecurity is the big bet. Booz Allen Hamilton says the model matched the recall of frontier cyber models in vulnerability detection with a favorable cost curve. Cloudflare highlights that defenders can scrutinize the architecture and run it on their own infrastructure with full sovereignty, and ENISA says open weights let organizations retain control over the technology their resilience depends on. The statements appear in Mistral's own announcement and were selected by Mistral.

For Danish companies, it is an obvious option if data needs to stay in Europe. The model is trained in Europe and can be hosted here. Special Danish availability does not appear in the sources.