Mistral releases open weights for its 1-trillion-parameter Mistral Large 4 model

Mistral AI released the open-weight version of Mistral Large 4, nicknamed Le Chonk, on 29 October 2026. According to the company's news page, the model entered public preview on Mistral Studio on 6 October; the preview announcement says it has 1 trillion parameters with 52 billion active, and was built from the ground up using 3,800 NVIDIA Grace Blackwell GPUs at Mistral's own data centers in Europe.

• The weights are on Hugging Face as Mistral-Large-4-1T-A52B and can be served with vLLM or SGLang.
• The license is the Mistral Ecosystem Open-Weight license, a modified MIT license: use, fine-tuning and commercial use are allowed, but companies selling model-as-a-service need a separate commercial license.
• According to Mistral's own results, the Terminal-Bench 4.0 score rose from 28.3 percent in the preview to 41 percent, and the model refuses malicious cyber requests 96.6 percent of the time on average.
• API pricing is 0.70 dollars per million input tokens and 2.10 dollars per million output tokens.

Why it matters

Open weights let organizations that do not want data to leave their premises run the model in their own data center or private cloud. That is an option worth assessing for banks, manufacturers and public bodies in Türkiye that are careful about personal data and trade secrets. Running a 1-trillion-parameter model in-house is likely to need substantial GPU capacity; smaller teams can also reach it through the Mistral API, OpenRouter, NVIDIA NIM and Microsoft Foundry. Mistral says its training data covers more than 160 languages; testing Turkish quality with a company's own samples may be wise.