Mistral Large 4: Open Weights, Operation, and API Pricing
Mistral Large 4, internally known by the somewhat less sober name “Le Chonk,” has not been available solely as an API preview since October 29, 2026. Mistral has released the weights of the improved version. The model processes text and images and has around 1 trillion parameters, 49 billion of which are active.
Mistral releases two official variants
Hugging Face offers two official variants of Mistral Large 4: FP8 and NVFP4. GGUF versions from Unsloth and the llama.cpp team are also available. The name sounds like a cuddly toy, but the model’s scale sounds more like a rather well-stocked data center.
The release is made under the Mistral Ecosystem Open-Weight License. This is a modified MIT license. It permits use, modification, fine-tuning, and commercialization of the model. However, companies offering Mistral Large 4 as a Model-as-a-Service require a separate commercial license.
So “Open Weight” should not be equated with “completely free” here. Those who use or adapt the model themselves have extensive freedoms. Additional conditions apply when operating it as their own Model-as-a-Service offering.
There are several ways to run it yourself
Hugging Face documents support for Mistral Large 4 in Transformers, vLLM, and llama.cpp starting with the release. For operation with vLLM, an NVFP4 variant on NVIDIA Blackwell GPUs is specified. A tested inference endpoint recipe uses eight NVIDIA H200 GPUs.
This is where “released” and “easy to start locally” diverge. The weights are available, but practical operation still depends on the appropriate software and sufficiently powerful hardware.
Those who do not want to operate the infrastructure themselves have several other access options. Mistral lists Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry, as well as NVIDIA-hosted Endpoints and NVIDIA-NIM. So the model does not just end up on the shelf as a download, but also in several hosted services.
Three reasoning levels are intended to control effort and results
Mistral Large 4 offers three reasoning levels. “Low” is intended for simple tasks, while “Medium” is for everyday tasks. “High” is set as the default for complex tasks.
The idea behind this is quite down-to-earth: Not every request needs to be handled with maximum reasoning effort. Anyone who only wants to complete a simple task should not automatically have to bring out the heavy computational machinery. However, the available information does not show how well this gradation works in everyday use.
Mistral reports 48 percent on the Artificial Analysis Coding Agent Index. At the same time, the model is said to require an average of 47 percent fewer tokens per task than the leading open comparison model. On Terminal-Bench 4.0, Large 4 achieves 41 percent according to Mistral. That is 13 percentage points more than the preview, with 30 percent fewer output tokens.
These figures come from the benchmarks cited by Mistral and accordingly represent the manufacturer’s claims or results from the tests mentioned. They provide points of reference, but do not replace your own evaluation using the tasks that actually arise in the respective project.
The API launches with clear token pricing
Since the release, Mistral has quoted $0.70 per million input tokens for the API. Cached input tokens cost $0.07 per million, while output tokens cost $2.10 per million.
This separates the pricing structure into input, cached input, and output. Especially for long tasks and extensive responses, the bill is therefore likely to reflect not just the model’s name, but also how many tokens flow in and back out. Le Chonk does not just have a big name.
Sources
See more of us in Google search results: add us as a preferred source.
Add as preferred source on GoogleWritten by
Daniel Böhmel has been developing software for over 20 years and has been self-employed since 2007. He started out with interactive web design back in the Flash era; later, he built his own online projects, such as a fitness forum and a supplement store, including the editorial content. Today, his…
View the profile of Daniel BöhmelParticipate now!
Don’t have an account yet? Register yourself now and be a part of our community!