Mistral's free AI model codes well and runs cheap, but China still leads

French AI company Mistral has released the downloadable files for Mistral Large 4, nicknamed "Le Chonk." It was trained entirely in Europe. Mistral says it beats other open models at coding. Early independent tests back that up in part: it is fast and fairly cheap, but it still trails the best Chinese models overall.

What it is

Large 4 takes text and images in and gives text and code back. "Open" here means the model's trained files, called weights, are public on Hugging Face, so companies can run it on their own servers. It is large, at about one trillion parameters. Running it yourself takes eight Nvidia B300 chips, or four GB300 chips for a more compressed version. Most firms will rent access instead.

How it scores

Mistral says extra training after the preview helped most with coding. On one coding test, Terminal-Bench 4.0, it rose from 28 to 41 percent, just ahead of GLM-5.3 at 40 percent. On another, FrontierCode 2.0, it leads open models at 19.7 percent. Claude Opus 5.5, a paid model, scores 46 percent there. In blind tests, people chose Large 4's code over that of GLM-5.3, Kimi K3 and DeepSeek more often than not. Against Opus 5.5 it won 22 percent.

On Artificial Analysis' overall Intelligence Index, Large 4 scores 26 points. That beats the median of 23 for open models of similar size. DeepSeek V4.1 Flash and GLM-5.3 score 32, and MiMo-V2.6-Pro scores 30.

Speed and price

• Through Mistral's service, it costs $0.70 per million input tokens and $2.10 per million output tokens.
• Tokens are the small chunks of text an AI reads and writes.
• It generates 161 tokens per second, up from 111 in the preview.
• An average index task costs $2.06, versus $5.01 for GLM-5.3 and $3.68 for Kimi K3.1 Pro.
• DeepSeek V4.1 Flash ($0.60) and MiMo-V2.6-Pro ($0.26) are much cheaper.

Security and licensing

Mistral calls Large 4 the leading open model for cyber defense, but its own chart shows two open models ahead of it. It refuses 96.6 percent of malicious cyber requests, more than any other open model. Mistral admits the model could still make real attacks easier and plans to tighten safeguards in the coming months.

The license is a modified MIT license. Use, changes and commercial deployment are free. Companies that sell Large 4 as a cloud service need a separate license from Mistral. It is also available on Microsoft Foundry.

What happens next

Mistral's lead among open models may be short. Moonshot plans to release the files for Kimi K3.1 Pro by November 4, 2026. It scores 67 percent on Terminal-Bench 4.0. Businesses weighing Large 4 should compare it with those rivals on their own tasks.