Mistral Large 4 Weights Are Out. The License Blocks Reselling It as a Service, and the Hardware Bar Is Eight GPUs. Use the API.
Follow-up to: 2026-10-23, when Mistral Large 4 was announced with the license unpublished. What's new: the weights were released on October 29, the license is public, and the API price is listed at $0.70 per million input tokens and $2.10 per million output tokens.
The headline for solo builders is that "open weights" and "I can run it" are different sentences. Mistral Large 4 is about one trillion parameters in total, and the hardware to serve it properly starts at eight GPUs.
What Mistral published
The model ships under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. Use, modification, fine-tuning, and commercial use are allowed. The catch is for businesses running a model-as-a-service offering, who need a separate commercial license from Mistral. Hugging Face's write-up calls it an MIT derivative that limits model-as-a-service offerings.
Mistral's API pricing is $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens. On benchmarks, Mistral reports 41% on Terminal-Bench 4.0, up 13 points from the preview, and 48% on the Artificial Analysis Coding Agent Index, roughly tied with GLM-5.3 while using 47% fewer tokens per task. These are Mistral's own figures. Hugging Face's post says that as of October 29 the model is the fifth best open-weights model on the Artificial Analysis Intelligence index.
The hardware reality
Hugging Face's guide gives the self-hosting picture. Serving in FP8 uses eight GPUs, and its Inference Endpoints setup uses eight NVIDIA H200s. The NVFP4 quantization lowers the memory footprint but needs NVIDIA Blackwell GPUs, and one example uses four B300s. Fine-tuning with LoRA in the TRL example uses a node of eight B300s.
The only single-machine route listed is the 2-bit GGUF quantizations, which can run on one Apple Silicon Ultra with 512 GB of unified memory. Heavier quantizations need distributed inference. A 2-bit version of a trillion-parameter model will behave differently from the full one, and the post doesn't publish quality numbers for it.
The context window is listed as 1 million tokens in the model overview, though the deployment examples use much smaller limits, such as 131,072 tokens in the vLLM example.
What the license means in practice
If you use the model inside your own product, say a coding tool or a support assistant, the license looks fine to me as written. If you wrap the weights in an endpoint and sell access to the model itself, that is the model-as-a-service case that needs a commercial deal. I'm not a lawyer, and where your product sits on that line is worth a written question to Mistral, not a guess from a blog post.
API versus your own box
For one person, the arithmetic is lopsided. I haven't priced eight H200s around the clock, and I'm not going to guess a figure, but I'd expect it to dwarf a typical solo product's AI budget. At $0.70 and $2.10, a million input tokens plus a million output tokens costs $2.80 through the API.
Self-hosting starts to make sense when one of these is true: a customer contract requires that data never leave your infrastructure, your volume is steady and high enough to keep the GPUs busy, or you need to fine-tune. Otherwise the open weights are valuable in a different way. They keep Mistral honest on price and give you a fallback if a vendor changes terms.
Where I could be wrong
Open weights have a second life through other providers. Third-party hosts can serve the model at their own prices, and some may undercut Mistral's API. I haven't checked who has listed it, and the license may limit some of them. Also, the 2-bit Mac route might be good enough for hobby use, and a 512 GB machine is a one-time cost, not a monthly rent. If you already own one, test it.
What I'd actually do
I'd call the API for a week with a sample of real prompts and compare it against the model I use now, on my own tasks, not Mistral's benchmarks. I'd keep the request format provider-neutral so swapping models is a config change. And I'd note the license text in my project docs now, so that if the product ever turns into a hosted model service, I find out before a customer does.
Author
Lukas
@lukcombinator