Mistral Large 4 weights arrive with a hosting license caveat
Mistral has released improved Large 4 weights following its October 6 public preview. The October 29 announcement opens a self-hosting path, but this is not an unrestricted MIT release. Teams evaluating it need to check the deployment package and business model, not just benchmark rankings.
What changed after the preview
Mistral reports 41% on Terminal-Bench 4.0, a 13 percentage point improvement over the preview. It also reports a 48% Artificial Analysis Coding Agent Index score with 47% fewer tokens than GLM-5.3 per task. These are attributed release claims, not ByteForward tests or a guarantee of lower total operating costs.
The public model card describes roughly one trillion parameters, with 49 billion active per token or 52 billion including embeddings and output layers. It accepts text and images, produces text, and supports a one-million-token context. Active parameter count should not be mistaken for the memory needed to store the whole checkpoint.
Which package should a deployment team inspect?
| Package | Observed artifact | Deployment distinction |
|---|---|---|
| FP8 | 1.06 TB repository | Start from the main model card |
| NVFP4 | 610 GB repository | Blackwell-oriented quantization, not a laptop promise |
Repository sizes come from the FP8 files listing and NVFP4 files listing. They are download footprints, not complete GPU memory requirements. No weights were downloaded for this article.
The NVFP4 instructions target B200, B300, GB200 and GB300 hardware. Their SGLang example uses tensor parallel size four, explicitly adjustable to the node topology. Their vLLM example instead uses eight and requires an FP8 KV-cache setting. Neither example establishes a universal four-GPU minimum or a guaranteed throughput figure.
Open weights do not remove the hosting restriction
The Mistral Ecosystem Open-Weight License permits modification and distribution with its notices retained. If a licensee or affiliate operates a Model-as-a-Service business, a separate commercial license is required before commercial use of the model or derivatives. The text excludes internal use and use through official Mistral products from that requirement. Its definition also distinguishes embedded end-user features and simple request relaying from Model-as-a-Service. Review the actual terms for your deployment rather than treating “open weights” as a complete licensing answer.
Reasoning settings and API prices
The main card documents low, medium and high reasoning effort, with high the default. Its request examples use temperature 1.0 and top_p 0.95. The completion-token budget includes both reasoning and visible output. For a first evaluation, record those settings alongside the exact checkpoint so comparisons remain reproducible.
The release announcement and model documentation now agree on $0.70 input, $0.07 cached input and $2.10 output per million tokens. These are hosted API rates, not the cost of operating your own GPU deployment.
A useful trial would compare the same small set of repository changes across reasoning modes. Check whether tests pass, inspect the changes, and log tokens, latency and actual charges. Include failed attempts in the totals. This is an illustrative evaluation plan, not a benchmark we ran. Lower token usage alone does not establish lower self-hosting costs.
AI-generated editorial illustration by ByteForward. Mistral icon reference from LobeHub Icons, © 2023 LobeHub, MIT License. Mistral is a trademark of Mistral AI. This independent illustration is not endorsed by Mistral AI. Ceramic objects are conceptual, not actual model hardware. Image reference license and modification notice.
