Mistral Large 4 Is a 1-Trillion-Parameter Model at $1.36 per Million Input Tokens, and the License Isn't Published Yet. Don't Rebuild Anything Until It Is.
Mistral announced Large 4 on October 6 as a public preview: a mixture-of-experts model with 1 trillion total parameters and 52 billion active, priced at $1.36 per million input tokens and $4.18 per million output tokens. The weights are due by the end of the month. What Mistral's announcement does not state is the license, and for anyone building a product on top of it, that one missing line matters more than every benchmark.
What was announced
According to Mistral's own page, Large 4 is natively multimodal and handles 160+ languages. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, and it is served on European infrastructure, including a deployment Mistral says it operates end to end. The company calls it the first milestone of its 3 billion euro Series D.
The page calls the model open-weight, says you will be able to self-deploy in a private cloud or on-premise once the weights ship, and then says nothing about under what terms. The context window is not stated either.
The numbers, with the right amount of salt
Every benchmark below is Mistral-reported. The page lists 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, 59.9% on AutomationBench and a 93% Cybench score. I have not found a third-party replication.
The one independent-looking number is a blind human coding evaluation by Surge AI, cited on Mistral's page: 3.74 out of 5, second of five models, behind Claude Opus 5 at 4.22. That is a good result for an open-weight model, and it also says plainly that the best closed model is still ahead by about half a point.
Mistral also says the preview's cybersecurity capabilities are offered to vetted partners and state authorities with reduced moderation during red-teaming. For a solo builder, the practical read is that you are getting the version with the guardrails on.
What it costs to call versus what it costs to run
At $1.36 per million input tokens, a single call is not cheap by 2026 small-model standards. OpenAI's GPT-6 Luna lists $0.10 per million input tokens, which makes Large 4 about 13.6 times pricier on input. They are not the same class of model, so that comparison tells you only that Large 4 is a premium tier you reach for on purpose.
Self-hosting is a different story. A 1-trillion-parameter model stores roughly 1 terabyte of weights at 8-bit precision and around 500 gigabytes at 4-bit, before you count memory for context. That is my arithmetic, and Mistral has not published inference hardware requirements. The practical conclusion is the same either way: nobody running a one-person business is serving this from a box under a desk. Someone else will host the weights, and then the question becomes whether they offer prices, regions and terms that are better than Mistral's own API.
Why the license matters more than the weights
"Open weights" has meant everything from permissive Apache-style terms to licenses that restrict commercial use above a revenue threshold or forbid certain uses. If the license caps commercial use at a size you might hit, or limits you from offering the model as a service, then a free download is a trial, not an asset.
That is why I would not change a line of code yet. Moving a production feature to a model whose terms you have not read is how a cost saving turns into a legal question.
Where this does help a solo builder
Vendor diversification. If your product depends on one US lab and that lab changes pricing, rate limits or terms, you currently have no equal alternative. A European-hosted, frontier-adjacent model with a published price gives you a second source you can test.
Data location. Some clients, especially in the EU, ask where prompts are processed. An end-to-end European deployment is a sales answer you did not have last month.
A benchmark to chase. Open models that land within half a point of the best closed model push prices down for everybody, which is good for you even if you never call Large 4.
Where I could be wrong
If the license turns out permissive and a few hosts offer the weights at a price below Mistral's API, Large 4 could become a default fallback model within weeks, and waiting costs you little. The other possibility is that real-world quality falls short of the vendor numbers, which is common in the first fortnight, and I would rather be late than early on that.
What I'd actually do
This week: nothing in production. Create an account on Mistral Studio, send it fifty real prompts from your product and compare answers and cost against your current model. That is an hour of work and it tells you more than the launch page.
When the weights land, read the license before the model card. Write down three answers: can I use it commercially at my revenue, can I offer it behind an API, and can I fine-tune and keep the result? If all three are yes, add Large 4 as a fallback behind your router. If any is no, you have saved yourself the migration.
Author
Lukas
@lukcombinator