Mistral Large 4 "Le Chonk" Goes Open Weight: What It Can Do and What It Takes to Run
Mistral put the weights for Mistral Large 4, nicknamed Le Chonk, on Hugging Face on October 29. It was one of the most upvoted posts on Hacker News that day. Here's a plain summary.
Size. About 1 trillion parameters in total, with roughly 49 billion active per token. It reads text and images and has a 1M-token context window.
Coding. It's one of the best open models here. It scores 41% on Terminal-Bench 4.0, up from 28% in the preview and just ahead of GLM-5.3 at 40%. Mistral says it ties GLM-5.3 on the Artificial Analysis Coding Agent Index (48%) while using 47% fewer tokens.
But. It's far behind the top closed model. Claude Opus 5.5 scores 46% on FrontierCode 2.0 versus Le Chonk's 19.7%, and in blind tests Le Chonk beat Opus only 22% of the time. On the overall Artificial Analysis index it gets 26, behind DeepSeek V4.1 Flash and GLM-5.3 at 32.
Security. Mistral calls it the top open model for cyber defense. The Decoder noticed Mistral's own chart puts it third among open models on the Cyber Index (51%, behind DeepSeek V4.1 Flash and MiMo-V2.6-Pro). It does have the highest refusal rate for malicious cyber requests among open models, 96.6% by Mistral's numbers.
Price and speed. $0.70 per million input tokens and $2.10 per million output through the API. Artificial Analysis clocked it at 161 tokens per second.
Running it yourself. The standard version needs eight Nvidia B300 GPUs, per SGLang's docs. So for most people it's an API, not something to run at home.
License. Free to use, modify and sell products with. Companies offering it as a hosted model service need a separate deal with Mistral.
My take: a solid release, especially for European companies that want a model they can audit and host. For a typical developer choosing on price and performance, the Chinese open models are still hard to beat.
Sources: Mistral, Hugging Face, The Decoder, Artificial Analysis, Hacker News.
Comments
Post a Comment