A downloadable checkpoint turns a promised deployment option into something a team can inspect.
November 3 update: Mistral released the weights of Mistral Large 4, also called Le Chonk, on October 29. The release announcement and model repository replace the earlier expectation that weights would arrive after the October 6 API preview.
The release package changes the deployment decision
The current model card describes a natively multimodal mixture-of-experts model with roughly one trillion total parameters. It lists 49 billion active parameters per token, or 52 billion when embeddings and output layers are included.
The card lists a one-million-token context window and text output from text and image input. It also describes low, medium, and high reasoning effort, with high as the default. These are Mistral’s specifications, not results from our deployment.
The license distinguishes several uses
The pinned Mistral Ecosystem Open-Weight License is a modified MIT license. If a licensee or affiliate operates a Model-as-a-Service business as defined there, commercial use needs a separate agreement with Mistral.
The definition concerns third-party access that gives meaningful control over inputs, parameters, or training data. The text excludes specified embedded end-user features and simple relaying of requests to models hosted by others. It also exempts internal use and use through Mistral’s official products.
Those distinctions should remain attached to any license summary. Downloadable weights do not mean unmodified MIT or Apache 2.0 terms. A business should assess its actual deployment against the full text before relying on an exception.
Active size and hosting memory answer different questions
The active parameter count concerns work selected for a token. It does not mean the rest of the checkpoint disappears from the serving system. Quantization, the inference stack, context state, and concurrent requests all affect the deployment budget.
The model card links serving guidance and an NVFP4 checkpoint. Treat that as a separate configuration with its own hardware requirements. An API invoice and a self-hosted deployment also have different cost structures; old preview sale prices should not be reused as a current hosting estimate.
Compare a fixed task before committing capacity
For a later evaluation, keep the repository snapshot, model revision, reasoning effort, and acceptance tests fixed. Record unsuccessful attempts and repair work alongside completed tasks. A preview result cannot be assumed to describe the released checkpoint.
We have not downloaded the weights, run the model, or reproduced Mistral’s coding and cybersecurity measurements. The release makes artifact inspection possible; workload performance remains a separate question.
Read the model card with the license. Which deployment constraint matters most for your project: hardware capacity, task quality, or permission to offer the model to others?



