French AI company Mistral AI released the full weights of its Large 4 model (ML4) on Oct. 29. For the first time, users can download the system and run it on their own hardware instead of accessing it only through Mistral's API.
Mistral said the model has roughly 1 trillion parameters, with 52 billion active at each inference step. It runs on eight Nvidia B300 GPUs, or four GB300 GPUs in a compressed version.
The weights carry a modified MIT license, though a separate Mistral license applies to commercial cloud deployment of ML4. The model is available on Hugging Face and Microsoft Foundry.
The company is marketing the release around cybersecurity: Mistral says the model refuses 96.6% of malicious requests tied to cyberattacks. CEO Arthur Mensch said the goal is for AI to be “a tool for empowerment, not dominance.”
According to benchmarks from Artificial Analysis, ML4 scored 48% on the Coding Agent Index, tying GLM-5.3 and trailing Claude Code's 62%. It scored 41% on Terminal-Bench 4.0, up from 28% for Mistral's previous model. On the Intelligence Index, ML4 scored 26 points, against 59 for leader Gemini 4 Carbon.
Via its API, the price is $0.70 for a million input tokens and $2.10 for a million output tokens, Mistral said, with a generation speed of 161 tokens per second. The company says it takes ML4 47% fewer tokens to complete the same task compared with GLM-5.3.
Mistral first announced Large 4 on Oct. 6. The weights were not downloadable until this week's release, three weeks later.
