# Mistral Large 4: what self-hosters need to know

> Mistral Large 4 weights are downloadable, bringing self-hosting choices and license conditions.

By BIGCHANGE.AI Editorial

Published: 2026-10-29T20:38:38.304Z
Updated: 2026-10-29T20:38:38.304Z
Canonical: https://bigchange.ai/blog/mistral-large-4-open-weights-self-hosting-license

![Conceptual charcoal illustration of an open rackmount server with fans and a finned module.](https://bigchange.ai/api/media/file/tensorrt-inference-server-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE; no actual NVIDIA product, build, measured route or accuracy result is depicted.

Mistral Large 4 weights follow its October hosted preview. The [repository](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/tree/main) lists files; the [release guide](https://huggingface.co/blog/mistral-large4) describes two formats. The choice has practical consequences: teams must provision the infrastructure and check a license that restricts model-as-a-service businesses.

## The big change

- **What changed:** Downloadable weights let developers run the model on their infrastructure and manage its operation.
- **Why it matters:** A smaller numerical format can reduce memory requirements, but format, GPU compatibility, usable context and simultaneous requests all affect the server configuration. Downloading weights is only the beginning of an operational decision.
- **What to watch:** Internal deployments and third-party model services face different license conditions. Organizations should check their intended use and test capacity against their own workloads before treating a published recipe as a fit.

## Deployment

Mistral’s [October 6 announcement](https://mistral.ai/news/mistral-large-4/) offered a preview API and said weights would follow by the end of the month. Our earlier [API guide](https://bigchange.ai/blog/mistral-large-4-api-guide) covered the hosted route. BIG CHANGE reviewed the linked documentation and repository listing, but did not download, run or benchmark the model.

The model card lists roughly one trillion parameters and 52 billion active per token, including embeddings and output layers. Its router selects experts for each token, but the server must store the model beyond those active parameters. It accepts text and images and produces text.

## Choosing a checkpoint and serving setup

Hugging Face identifies FP8 and NVFP4 as the two official variants. Quantization represents parts of a model with lower-precision numbers, reducing storage and memory demands compared with a higher-precision representation. The [NVFP4 card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVFP4) targets NVIDIA Blackwell GPUs and combines NVFP4 experts with FP8 linear layers. Mistral reports similar evaluation results to FP8; that is not a guarantee of identical results on every task.

The main repository listing reports 1.06 TB. Disk size is not a complete GPU-memory requirement: serving also needs space for runtime state and requests. Hugging Face describes an FP8 endpoint recipe using eight H200 GPUs, with a 131,000-token context limit and up to 16 concurrent requests. This is an example configuration, not a universal minimum. Its guide also shows a four-B300 deployment example for NVFP4.

The card states a one-million-token context window; a particular server may be configured for less. Longer inputs and more simultaneous requests consume memory. Operators should check current serving instructions for their hardware and chosen checkpoint rather than assume a generic install command establishes capacity.

## What the license permits—and where to check

The [Mistral Ecosystem Open-Weight License](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) is a modified MIT license, not unrestricted MIT. It permits use, modification and distribution subject to its conditions, including retaining the required notices. Its model-as-a-service clause requires a separate commercial license before commercial use by a licensee or affiliate operating such a business; it defines that service in terms of third-party access and meaningful control over inputs, parameters or training data.

The text excludes internal use and use through Mistral’s official products from that clause, and distinguishes embedded product features and merely relaying requests to models hosted by others. The exact terms and a team’s actual service design matter, especially for offerings that sit between an internal application and a general-purpose endpoint. Read the license itself and obtain appropriate legal advice for the intended deployment.

## Sources

- [Mistral October 6 announcement](https://mistral.ai/news/mistral-large-4/) — Hosted preview and planned weights.
- [Mistral model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B) — Model architecture and context.
- [Mistral model files](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/tree/main) — Checkpoint listing.
- [Mistral NVFP4 card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVFP4) — Precision and hardware.
- [Hugging Face guide](https://huggingface.co/blog/mistral-large4) — Formats and deployment examples.
- [Mistral license](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) — License text and service condition.