Mistral released Large 4 weights: about 1 trillion parameters, $0.70 in and $2.10 out per million tokens, and eight B300 GPUs to self host

Open weights does not mean you will run it in the office, but it still matters for some clients.

What happened

Mistral published the weights for Mistral Large 4, nicknamed "Le Chonk," on Hugging Face. The Decoder reports it has about one trillion parameters with 52 billion active per step, takes text and images, and outputs text and code. According to SGLang's docs cited by The Decoder, the standard version runs on eight Nvidia B300 GPUs.

The license is a modified MIT license: use, modification and commercial deployment are free, but offering Large 4 as a commercial cloud service needs a separate license. API pricing is $0.70 per million input tokens and $2.10 per million output. It is also on Microsoft Foundry.

On performance, Artificial Analysis rates it 26 on its Intelligence Index, behind DeepSeek V4.1 Flash and GLM-5.3 at 32. It generates 161 tokens per second, and an average index task costs $2.06, against $0.60 for DeepSeek V4.1 Flash. On coding agent tasks Mistral says it roughly ties GLM-5.3 while using 47% fewer tokens. TechCrunch quotes Mistral's Pierre Stock saying open weights give access "that can't be cut off."

My take

For most small business automations, this is not a self hosting story. Eight top end GPUs is not a side project. The realistic route is the API or a cloud platform.

Where I would actually consider it:

1. European clients with data residency questions. A model trained in Europe that can run on infrastructure you choose is an easier conversation with compliance.
2. A fallback model. If a provider changes terms or pricing, an open weight option you can move elsewhere is insurance.
3. Speed sensitive steps. 161 tokens per second helps chat and voice replies.

For simple extraction and classification, cheaper models still win on cost per task. Test on your own workflows before switching, and keep the model name as a setting, not something hard coded into every step.