Mistral Releases Large 4 With Open Weights: 1T Parameters, 52B Active

French company Mistral has released open weights for Large 4 on Hugging Face, judging by a post on X. The model has 1T parameters, with an architecture that activates 52B of them for each token. The context window is 1M tokens. The license is a modified MIT with a separate exception for model-as-a-service. The post's author also cites API pricing and a 63% reduction in the cost of completing intelligence-index tasks. One important caveat: the source is a single post with tiny reach, and the text links to neither an official announcement nor a model card. All figures below are therefore the author's claims, not verified data.

What is claimed about architecture and context

The ratio of 52B active parameters to 1T total means roughly 5% of the weights are used per token. This sparse approach (mixture-of-experts) lets a model carry enormous capacity while inference costs stay closer to those of a much smaller dense network. For anyone hosting the model themselves, it does not remove the main constraint: all the weights, that is, a trillion parameters, still have to sit in memory. Even at low precision, that means a sizable cluster, not a single workstation.

A 1M-token context is a pitch for working with large codebases, long documents and multi-step agent workflows. The post says nothing about how quality holds up over long distances. A stated window size and the real ability to find and use information at its edges are different things, and independent tests need to check the latter.

License: MIT with a caveat

The author calls the license a modified MIT with an exception for model-as-a-service. The post does not spell out the exact wording. From the name, one can guess that the exception restricts third parties from commercially hosting the model itself as a service, but that is a guess, not a fact. Companies planning to use Large 4 inside their own products probably face few restrictions. Inference providers and API resellers should read the original license text first. Until it has been reviewed, it is too early to call Large 4 fully open in the spirit of a pure MIT license.

Pricing and task economics

The author states that the cost of completing one intelligence-index task fell 63% and now stands at $2.06. The post does not name the index or the model used for comparison. Taken literally, the previous cost was about $5.57 per task — that is our calculation from the stated percentage, not the author's figure. A drop in task cost depends on both token prices and how many tokens the model spends on reasoning: a cheap rate does not help if the answers come out verbose.

API rates are $0.70 per million input tokens and $2.10 per million output tokens. Output costs exactly three times as much as input. For scenarios with long context and short answers, such as document search or code review, this structure works in the user's favor. For generating large volumes of text or long reasoning chains, output tokens will drive most of the bill.

What to check before adoption

A minimal checklist for a team considering Large 4:

• compare the post against Mistral's official announcement and the model card on Hugging Face;
• read the full license text and the terms of the model-as-a-service exception;
• find out which index and which baseline model lie behind the 63% cost reduction;
• run your own long-context tasks instead of relying on the stated 1M;
• estimate the memory requirements for self-hosting a trillion-parameter model.

If the claims hold up, Large 4 would become one of the largest open-weights models that has both public weights and an inexpensive API. For now, this is a report with no documentation as its source. It makes sense to treat it as a reason to verify the information, not as grounds to change a product's architecture.