Mistral Large 4 is a 1-trillion-parameter mixture-of-experts model from Mistral AI, and at the end of October the company put its weights on Hugging Face. Only 49 billion parameters are active per token (52 billion with embeddings), the checkpoint weighs 1,061 GB, and the Mistral Large 4 license is a modified MIT with one extra condition about selling access to the model.
In this post I go over three questions for a five-person team: can we put it in a product, can we host it, and is the API cheaper. My verdict: at Trellis’s size this is a “watch” model, not a “switch” model. Keep a model-swap layer in your code and pay the API until your volume says otherwise.
I did not run this model, because there is no terabyte of GPU memory in my sandbox. I read the LICENSE file, the config.json and the file index on Hugging Face, then did the maths. Mettons that makes me a careful reader, not a tester.
What does the Mistral Large 4 license actually allow?
The license file is called the “Mistral Ecosystem Open-Weight License”, and it opens by saying it is “a modified version of the MIT license”. You may use, copy, modify, distribute, sublicense and sell the model, deploy it and train it further. Then comes Section 2:
If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.
So the whole question is what “Model as a Service” means. The file defines it as “giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints)”, in a way that lets that third party “exercise meaningful control over the inputs, parameters, or training data.”
Then the carve-out, which is the sentence a startup should read twice. The definition “does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.” Section 3 adds that internal use, anything that does not make the model available to third parties, is outside Section 2 too.
For a SaaS like Trellis, where the model sits behind a feature and a seller never writes their own prompt against it, that carve-out looks friendly. Looks. I am not a lawyer, and “meaningful control” is doing a lot of work. If your product lets customers bring their own prompts or fine-tune on their data, you are closer to the line than you think, so take Section 2 to your counsel before you build that. Hugging Face also lists the license as “other”: it is not OSI-approved.
Can a small team host Mistral Large 4?
Short answer: not on the GPUs most of us can rent. The checkpoint is fp8 in 54 safetensors shards, and the sizes add up to 1,060,548,427,328 bytes, the same number as the index file’s total_size. The model card example for vLLM uses --tensor-parallel-size 8, so the vendor is thinking in nodes of eight GPUs.
Here is my script’s output, trimmed, with arrows changed to words so WordPress keeps them.
fp8 checkpoint bytes 1060548427328 = 1061 GB
8xH100 80GB total VRAM 640 weights DO NOT fit at fp8 headroom -421 GB
8xH200 141GB total VRAM 1128 GB; weights fit headroom 67 GB
8xB200 180GB total VRAM 1440 GB; weights fit headroom 379 GB
ctx 1048576: latent cache per sequence bf16 = 70.9 GiB (estimate)
blended $/M tokens at 5:1 in:out = 0.933
node $20/h x730h = $14,600/mo so break-even 15.6B tokens/month (0.52B/day)
node $40/h x730h = $29,200/mo so break-even 31.3B tokens/month (1.04B/day)
node $60/h x730h = $43,800/mo so break-even 46.9B tokens/month (1.56B/day)Eight 80 GB cards give you 640 GB, and the weights alone need 1,061. They do not fit. Eight 141 GB cards give 1,128 GB and fit with about 67 GB to spare, which is thin, because the spare room is where your KV cache lives.
The two cache lines are my estimate, not a measurement: 63 of the 72 layers in config.json own a cache, at 576 latent dimensions per token, assuming a latent-style cache in bf16. Engines differ, so treat it as an order of magnitude. A 1M-token context is not free.
Quantized builds exist for smaller hardware (an NVFP4 one around 610 GB for Blackwell cards, and GGUFs down to about 298 GB), and that is my one sentence on them.
Is the Mistral Large 4 API cheaper than hosting it?
Almost certainly yes, at our size. I re-fetched Mistral’s pricing page on October 30 and the row for Mistral Large 4 reads $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. The page shows no effective date, so look again before you budget.
Assume a 5-to-1 mix of input to output tokens, which gives a blended API price of $0.933 per million. Now take a node of eight big GPUs running all month, 730 hours. I do not know your quote. The last four lines of the output above use three hourly prices ($20, $40 and $60) that I made up, so plug in your own.
The formula is the useful part: node cost per month divided by the blended price gives the monthly volume where hosting starts to win. It ignores the second node and the weeks of tuning.
Now a made-up Trellis day. Say 2,000 requests at 6,000 tokens each: 360 million tokens a month, about $336 in API fees at the blended price. The cheapest made-up node breaks even at 15.6 billion, about 43 times more. Hosting a trillion-parameter model for that is a bus for two people.
Cached input at $0.07 helps if your prompts share a prefix. I did not model it, so $336 is probably high.
What are the critics saying about Mistral Large 4?
Two arguments are worth your time. The first is about the word “open”. In the Hacker News discussion, rcr-anti wrote that this is “weights available”, not open weight, because the license is a modified MIT that is not OSI-approved, like BSL or SSPL. Taek disagreed, saying most people call anything with downloadable weights “open weight” and nobody is surprised by the restriction.
Commenter dannyw read the internal-use clause and concluded you can host and run it freely, and only need a license to sell access as an inference provider. Franchement, I side with the strict readers on the label and the practical ones on the effect: a startup that is not reselling inference is probably fine. Probably.
The second is about quality. Mistral’s own announcement claims 41% on Terminal-Bench 4.0, 84.5% on BrowseComp and a 48% score on the Artificial Analysis Coding Agent Index, which it calls “effectively tied” with GLM-5.3. Those are Mistral’s claims, not mine. Commenter oh_no called the benchmark numbers “dreadful”, and almindor, who uses it on a personal project, found it too expensive and very slow when thinking.
One commenter, sigmar, argued the weights may lack copyright, so the terms behave like a website’s terms of service. That is a theory. Do not build a company on it.
Watch or switch: what I would do on Monday
Do not host this, and do not rebuild on it. Keep one thin layer between your product and the model, so that changing providers is a config change and a regression run, not a rewrite. The Mistral Large 4 license can change, and so can the price. The Muse Spark license post I wrote a couple of weeks ago reached the same place with a different license: the value of published weights is the option to move later.
I would change my mind in two cases. If your customers need data to stay on your own hardware or in the EU, this is one of the few models of its size you can put there. If your volume is in the billions of tokens a month, run your own node quote through the formula. And if customers write their own prompts on your platform, use Mistral’s contact form first, because that is the case Section 2 is written for.
What I will do next is the one test I can do without a terabyte of memory: send twenty Trellis-style prompts to the API with the model name in a config file, and count the lines I had to touch to swap it.
