November 5, 2026 · Open Weights
A trillion-parameter model you can download, and the clause that decides who may use it
On 29 October Mistral put the weights of a trillion-parameter model on Hugging Face. The interesting question for a cross-border team is not how it scores. It is what the licence lets you sell.
On 29 October Mistral released the weights of Mistral Large 4, the model it had previewed earlier in the month under the nickname Le Chonk. The headline numbers: about 1.05 trillion parameters in total with roughly 50 billion active per token, native text and image input, a one-million-token context window, and support for more than 160 languages. Mistral says it was trained on around 3,800 Grace Blackwell GPUs inside its own European data centres.
For a cross-border team, “open weights” has usually meant a model small enough to run and too weak to matter. A trillion-parameter release moves the arithmetic in two directions at once. You can now put a frontier-adjacent multilingual model behind your own firewall. You cannot put it on the server you already rent.
What the benchmarks say, and what they do not
Mistral reports 48% on Artificial Analysis’s coding-agent index, which it describes as roughly level with the leading open-weight model while using about 47% fewer tokens per task. Terminal-Bench 4.0 comes in at 41%, up 13 points from the preview. On Cognition’s FrontierCode 2.0 — which asks whether a maintainer would merge the agent’s change — it scores 19.7%, first among open-weight models. Defensive security work lands at 51% on the AA Cyber Index, with a stated refusal rate of 96.6% on malicious cyber requests.
Two caveats belong next to those numbers. They are vendor-reported, and the most interesting one — tokens per task — is also the least comparable across labs, because it depends on how the harness is configured. A model that finishes the same task on half the tokens is a better deal for an agent that runs all day. It is not automatically a better model.
The licence is the story
Large 3 shipped under Apache 2.0. Large 4 ships under what Mistral calls an ecosystem open-weight licence, a modified MIT: free to use, modify, fine-tune and commercialise, with one carve-out. If you sell model-as-a-service — if the model is the thing your customer is paying for — you need a separate commercial agreement.
That carve-out is where the industry is drawing its line, and it is being drawn around inference resellers rather than around users. Running the model inside your own company is fine. Embedding it in a product where the model is incidental to the value is, on a plain reading, fine. Standing up an API and selling tokens is not. If your roadmap includes the third case, the cost of the licence belongs in the model before the cost of the GPUs.
Where this lands on an export desk
The first use is language. One deployment covering 160-plus languages is a different proposition from stitching together four regional vendors, particularly for the languages outside the big three — Polish, Thai, Vietnamese, Turkish. Product copy, marketplace localisation, supplier correspondence and after-sales scripts can share one model and one glossary, which is the part that actually protects brand consistency.
The second is residency. “Where is our product data processed?” is now a routine procurement question from European buyers, and running the model on your own hardware answers it in a way a hosted API cannot. For teams selling into the EU, that answer is sometimes worth more than two points of benchmark.
The third is contract review. A million-token context means a framework agreement plus its annexes fits in a single pass without chunking, and chunking is precisely where document pipelines quietly lose clause references.
The hardware reality
A trillion parameters does not compress into good intentions. Mistral’s own example configuration is eight H200-class accelerators; even aggressive quantisation lands in the very-large-unified-memory bracket rather than the spare-server bracket. Add power, cooling, redundancy and the person who maintains it, and self-hosting becomes a three-year commitment rather than a monthly line item.
Which produces a simple test. Take your monthly spend on hosted inference for the same workload and multiply it by 36. If that number is smaller than the hardware, the rack and the engineer, self-hosting is a hobby. If it is larger — and stays larger once you account for utilisation, because a privately hosted model idles all night — the calculation flips.
What to do now
- Read the licence before the benchmark. Model-as-a-service is defined by what your customer pays for, not by what you call your product.
- Run a 200-item bake-off on your own content in your own languages, before you believe any index.
- Measure tokens per finished task, not tokens per call. That is the number your invoice follows.
- Keep a hosted fallback. The open-weight tier is moving weekly; locking in today’s leader is a bet, not a strategy.
- Check residency end to end. Weights on your own metal is half the story if the tokeniser and evaluation harness still phone home.
The significance of this release is not that a European lab caught up on a leaderboard. It is that the licence now separates “can I use this” from “can I resell this”, and that distinction will decide more business models over the next year than any two points of benchmark.