I read the license before I read the benchmarks, which isn’t how I usually spend a release day. Mistral put the weights for Mistral Large 4 up on October 29, three weeks after the October 6 preview, under a license nobody had seen before the repo went live. The headline everyone repeated is that this is the best open model outside China. The part that decides whether you can actually ship on it is Section 2, and Section 2 has no number in it.
🐈 What actually shipped
The model is real, and it’s big: about 1.05 trillion total parameters, 49 billion active per token (52 billion counting embeddings and output layers, a definitional difference and not a quiet model swap during the preview, as VentureBeat checked). Architecturally it’s a sparse mixture of experts with 128 routed experts, four active per token, plus one shared expert, a 1M token context window, and image input through a Pixtral style vision encoder. Training ran on 3,800 Grace Blackwell GPUs in its own European datacenters, and the sovereignty pitch runs through the whole announcement.
Two official checkpoints went up, FP8 and NVFP4, plus GGUFs from Unsloth and llama.cpp. I counted the shards, because the download size tells you who this release is for. The FP8 repo is 54 safetensors files at about 20 GB each, so roughly 1.1 TB. The NVFP4 repo is somewhere around 600 GB. Mistral’s own vLLM recipe asks for --tensor-parallel-size 8, which is eight high end GPUs before you allocate a single byte of KV cache.
API pricing dropped to $0.70 input, $0.07 cached input, $2.10 output per million tokens, roughly half the preview rates.
The 63% lower cost per task that VentureBeat reported from Artificial Analysis is a real gain. That isn’t the story.
⚖️ Section 2, read out loud
Here is the operative text, from the LICENSE file in the repo:
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.
Here’s the definition that decides who that covers:
“Model as a Service” means 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 manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This 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 carves out internal use that doesn’t make the model available to third parties, and use through Mistral’s own products.
Read it fairly and the scope is narrow. Internal deployment is fine, even for a company that also runs an inference business, and VentureBeat’s reading of the text lands in the same place. Relaying requests to someone else’s Mistral endpoint is fine. Shipping a product where the model sits inside a feature is fine.
The trouble is the words doing the load bearing: “solely”, “meaningful”, “harnesses”. None are defined, there are no examples, and there is no safe harbor short of emailing the contact form.
🧾 The threshold got replaced, not removed
Mistral Large 3 shipped under Apache 2.0. Some of Mistral’s other open releases use a modified MIT that the company’s own help center describes like this: same permissions as Apache 2.0, except that companies with monthly revenue above $20M either buy a commercial license or use Mistral Studio. That page is still up today and still uses Devstral 2 as its example. It doesn’t mention a model-as-a-service clause anywhere.
The new license, the Mistral Ecosystem Open-Weight License, is also a modified MIT. It keeps the permissive grant and the attribution requirement. What it swaps out is the test. Twenty million dollars a month is a number you check against your books. “Operates a Model as a Service business” is a judgment about what your product is.
Hugging Face’s metadata for the repo says license: other, which means the platform doesn’t recognize it as a standard license. Any compliance scanner in your pipeline will flag it for manual review, and the reviewer will be reading the same paragraphs you are.
I’ll say the thing Mistral’s marketing avoids: a field-of-use restriction means this isn’t open source. The OSI definition forbids discriminating against fields of endeavor, and “you may not run a model-as-a-service business with this” is exactly that. Open weights is a fine term. Open source is a different one, and only one is accurate here.
🚧 Where the line actually falls
The interesting failure mode isn’t the person who obviously operates an inference provider. It’s the person in the middle, which is most of us.
Suppose you host the model and let customers send prompts. Under the definition, you’re giving a third party access for inference, and they control the inputs, because the prompt is theirs. Under exclusion (a), you’re an end-user product with model capabilities embedded in a feature or a harness. Both descriptions fit the same architecture diagram. If you also let the customer pick a reasoning effort, or set temperature, you’ve arguably handed them control over “parameters”, which is a decoding parameter in the ordinary engineering sense and a loaded word in the license.
Fine tuning is clearer: exposing post-training to customers is squarely inside the definition, so a tuning platform needs the commercial agreement.
The affiliate clause is where it gets uncomfortable for anyone with a hosting arm. The trigger isn’t only that you run a model-as-a-service business. It’s that you or any affiliate does. Once that’s true, the license requires a separate agreement before using the model “for any commercial purpose”, which is broader language than the definition it’s attached to. Internal use is carved out. A commercial product is not obviously carved out just because it falls under exclusion (a), because that exclusion narrows the definition of model-as-a-service, not the scope of the obligation. I might be reading it too hard. That’s the problem: I can’t tell, Mistral hasn’t published guidance, and a forty word clarification would settle it.
What the license doesn’t have is just as notable: no revenue number, no user count, no audit right, no termination, no governing law, no definitions beyond the paragraph quoted above. For a permissive license most of that is harmless. For the one clause with teeth, it means the teeth point at whoever has the least appetite for a legal opinion.
📊 The benchmarks come with asterisks
Mistral’s own materials are more careful than the coverage. The model card says 26 on Artificial Analysis Intelligence Index v5 and calls it the highest among western open-weight models, while noting directly that the leading Chinese open models score higher. That matches what Artificial Analysis published on October 29: 26, up from 25 at preview, and six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3, both at max reasoning effort.
On the Coding Agent Index, Artificial Analysis has it at 48 with Mistral Vibe, one point behind GLM-5.3 at 49 running in Opencode, and two above DeepSeek at 46. Mistral calls that effectively tied, Artificial Analysis ranks it second. The win that survives is token efficiency: 8.5 million tokens per task against 15.9 million for GLM. That’s the number I’d care about when sizing an agent budget.
Terminal-Bench 4.0 is where the vendor figure and the independent figure part company. Mistral reports 41%, a 13 point gain over preview and the lead among open weights. Vals AI, running the released OSS checkpoint at high reasoning effort, reports 26.26%, eighth of 22 open-weight models and 26th of 58 overall. Same benchmark name, different harness and provider, so these aren’t the same measurement. But a 15 point gap between the launch number and an independent run of the same checkpoint is exactly why you run your own eval instead of adopting the headline. Vals also puts the model 17th of 30 on its own index at 56.42%, fifth among open weights.
Cost per Intelligence Index task came down to $2.06 from $5.54. DeepSeek is at $0.60 and MiMo-V2.6-Pro at $0.26 on the same measure. Good improvement, not the cheapest seat in the room.
🧠 Final Thoughts
The whole pitch for this release is control. You get the weights, you host them yourself, nobody can disconnect you mid incident, and the training lineage is European end to end. I think that pitch is legitimate, and the refusal behavior on defensive security tasks is a real reason to prefer a model you own over one behind a policy filter.
But the license draws the boundary of that control exactly where the hardware does. Almost nobody who can afford eight Blackwell GPUs is doing it to run a chatbot for their own team. They’re doing it to serve someone else, and serving someone else is the thing that requires the phone call. The people the license exempts are the ones with a data center and no customers. The people it captures are the ones with customers.
If you’re evaluating Large 4 this week, decide what you are before you download 1.1 TB. Internal tooling, go ahead. Fine tuning as a service, get the agreement first. A product with the model embedded in a feature, write down your reasoning and keep it somewhere your future self can find it, because the terms won’t get clearer and the ambiguity doesn’t expire.
Thanks for reading.