Mistral releases Large 4 weights with commercial conditions for model hosts

Mistral has released the final weights for Large 4, turning its October preview into a model that organisations can download and operate themselves. The company says it used the intervening three weeks for further reinforcement learning, so the release is also a new checkpoint rather than simply a download of the original preview.[1]

The distinction matters to buyers who have been comparing access routes. On 6 October, Mistral offered Large 4 through an API and promised weights by the end of the month. The new release fulfils that promise; the practical questions now concern deployment resources and the exact permissions attached to those weights.[2]

Two parameter counts, one model

The repository names the model A52B, yet its description also says 49 billion parameters are active per token. The model card reconciles the figures: 49 billion covers the active computation before embeddings and output layers; including those brings the count to 52 billion. The complete mixture-of-experts model has roughly one trillion parameters.[3]

That separation is useful when interpreting deployment claims. Activating only part of a model for each token reduces computation, but the active count alone cannot tell an operator how much memory is needed to hold the full checkpoint, serve concurrent users and retain their context.

The card describes text and image inputs, text output and a one-million-token context capacity. It provides serving instructions for vLLM and SGLang and points to a separate NVFP4 checkpoint targeting Blackwell GPUs. Those are concrete deployment routes; the supported context length and the capacity affordable on a particular installation remain different quantities.[3]

The licence follows the business

The Mistral Ecosystem Open-Weight License permits modification, deployment and further training, subject to its conditions. Clause 2 requires a separate commercial licence if the operator or any affiliate runs a Model-as-a-Service business as defined in the licence, before using the model or its derivatives for any commercial purpose.[4]

That definition concerns third-party access to Large 4 or its derivatives for inference or post-training, with meaningful control over inputs, parameters or training data. It excludes capabilities embedded solely in specific end-user features or harnesses, and the mere relay of requests to models hosted elsewhere. Clause 3 exempts internal use that provides no third-party model access, and use through Mistral’s official products.[4]

For a company offering both software products and model hosting, the clauses and exceptions need to be read across the organisation’s activities. The product interface alone gives an incomplete picture.[4]

A price reference for the deployment decision

At those rates, one million uncached input tokens plus one million output tokens would cost $2.80 through the API, before any other applicable charges. Self-hosting replaces that token bill with hardware, energy and operating costs. Utilisation and concurrency therefore matter alongside the model’s performance on the buyer’s own tasks.[5]

Report an error or suggest a correction