Open Weight Models: Mistral Large 4 Has No Revenue Cap, but Its License Turns on Undefined Words

Mistral published the weights of Mistral Large 4 this week. The checkpoint is a mixture-of-experts model of about one trillion parameters, with text and image input and a 1M-token context. Whether you are allowed to put it behind your own product is a different question, and the answer is in the license file.

What shipped

The config file lists 128 routed experts plus one shared expert, four experts active per token, 72 text decoder layers, a hidden size of 6144, and a maximum position count of 1,048,576. The FP8 checkpoint stores weights in 128 by 128 blocks and leaves the vision tower, the output head and a few attention projections unconverted. Mistral also lists an NVFP4 variant, and Hugging Face lists GGUF builds from ggml-org and Unsloth.

The announcement page does not state the active parameter count. The Hugging Face post says 49B, while the repository name says A52B. The model card settles it: 52B includes the embedding and output layers. So the two figures are the same model counted two ways, and 52B is the number to use when you size memory bandwidth.

Mistral's benchmark claims are mostly its own or come from vendor-run evaluations, and I did not reproduce any of them. One of the headline results is on a private benchmark run by Cognition. Treat them as a reason to test, not as a result.

The hardware floor for self hosted open weight models

Nothing on the announcement or the model card states a minimum. The examples tell you the floor. The vLLM recipe uses tensor parallelism of 8, and the NVFP4 checkpoint needs Blackwell GPUs. Hugging Face's Inference Endpoints recipe uses 8 H200s. The llama.cpp route is the single-machine path the post describes: the Hugging Face post says 2-bit quants run on one Apple Silicon Ultra with 512 GB of unified memory, and the author ran Q2_K on an M3 Ultra.

Simple arithmetic on the config shows why: a trillion parameters at one byte each is on the order of a terabyte of FP8 weights before any cache. If you plan around the 512 GB route, run your own evaluation on that exact quant.

The license, read from the file

The model card calls it a modified MIT license, and the LICENSE file confirms the shape. Use, modification, fine-tuning, distribution and commercial use are granted. There is no revenue threshold anywhere in it. That is the difference from the Timestral license covered in our earlier teardown, where a monthly revenue figure withdrew all rights.

The restriction is narrower and vaguer. If you or an affiliate give a third party access to the model or a derivative for inference or post-training, including through an API or hosted endpoint, you need a separate commercial license from Mistral. The test in the file is whether that third party gets meaningful control over the inputs, parameters or training data. Internal use is exempt. Features embedded in a specific product or harness are excluded, and so is relaying requests to a model someone else hosts.

The file does not define meaningful control. A chatbot with a fixed system prompt probably sits on the embedded side. A fine-tuning service sits on the other. An internal platform team exposing a shared endpoint to ten business units is internal use. The same endpoint exposed to a subsidiary or a customer is where the file stops helping.

The claim, and what would break it

My claim: for most teams the license is not the blocker, but for anyone whose product lets customers send arbitrary prompts or choose sampling settings, it is the first item in the plan, ahead of GPUs. A cost model that ignores it is incomplete.

This is wrong if Mistral publishes a definition or a worked FAQ that places typical customer-facing assistants clearly outside the clause. The license points to a contact form for edge cases, and I found no such document on the pages I read. It is also wrong if your product only exposes fixed features and never raw model parameters.

What to do this quarter

Mistral's hosted price is $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens, according to its announcement. That sets the bar a self-hosted deployment has to beat. Run this before ordering hardware:

1. Pull one month of your token counts and compute (input / 1e6 x 0.70) + (output / 1e6 x 2.10). If that number is smaller than the monthly cost of an eight-GPU node, hosting yourself is a data-control decision, not a savings one.
2. Answer in writing whether any third party, including a customer, an affiliate or a partner, sends the model inputs or sets parameters. If yes, open the commercial license conversation before the pilot, not after.
3. Run your own evaluation set against the exact quantization you intend to serve, and keep the FP8 or API result as the reference. A gap above the tolerance you set in advance means the quant is out.
4. Pin the vLLM image named in the model card rather than a floating tag. The card names a model-specific image, vllm/vllm-openai:mistral-large-4.

The broader pattern is the one in why enterprise AI does not ship: the model was never the hard part. Here the model is free to download, and the last mile is a legal definition and an eight-GPU node. If your private deployment plan is to move off an API for privacy reasons, our private LLM deployment service starts from exactly those two constraints.