---
title: "Mistral releases Large 4 weights, bringing deployment control and licensing trade-offs"
description: "Mistral releases Large 4 weights on October 29. Explore verified availability, benchmark results, hardware requirements and model-as-a-service license limits."
url: "https://promptszone.com/mistral-large-4-open-weight-release"
published: "2026-10-29T16:00:06Z"
categories: ["New Models"]
---

# Mistral releases Large 4 weights, bringing deployment control and licensing trade-offs

Mistral's October 29 release makes Large 4 weights downloadable. Coding results, infrastructure requirements and a model-as-a-service license restriction shape the trade-offs.

![Original Mistral Large 4 scorecard: independently verified Artificial Analysis Intelligence Index v5.0 score 26 for high reasoning; Mistral-reported Coding Agent Index 48%. Different metrics, checked October 29, 2026.](https://promptszone.com/assets/uploads/f342c48b0800161c4395cf09d6091aed.png)

**Summary:**

Large 4 moves beyond its October 6 API preview with downloadable weights. Its sourced scorecard separates independent intelligence testing from developer-reported coding results.

Mistral has released downloadable weights for Mistral Large 4 on October 29, moving its flagship beyond the API preview introduced on October 6. The release gives organizations a practical route to operating and adapting the model themselves, although its infrastructure requirements and commercial license deserve as much attention as its benchmark results.

The change is verifiable in the [official repository](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/tree/main): it now contains model-weight shards, rather than a placeholder for a future release. The main repository lists roughly 1.06 TB of files. That is a download footprint, not a complete estimate of the memory or hardware needed to serve it.

## What is available now

The [model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B) describes a roughly one-trillion-parameter mixture-of-experts model with text and image inputs, text output and a one-million-token context window. It distinguishes 49 billion active parameters per token from 52 billion when embeddings and output layers are included. Three reasoning settings let applications choose how much work the model spends on a request.

Mistral also supplies an [NVFP4 checkpoint](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVFP4) intended to reduce memory use. Its documentation targets NVIDIA Blackwell hardware and provides serving recipes. That makes the release more accessible to appropriately equipped operators, but it should not be read as a promise that the flagship will run economically on a typical personal computer. Quantization, concurrency, context length and available accelerator memory all affect the deployment calculation.

## A coding story, not a universal benchmark victory

In its [release announcement](https://mistral.ai/news/mistral-large-4-oss/), Mistral reports 48% on the Artificial Analysis Coding Agent Index and says the model uses fewer tokens than a leading open-weight comparator. The company also reports improvements since the preview on terminal-based coding tasks. Those are useful leads for evaluation, but neither a token reduction nor a benchmark position establishes the cost or reliability of a particular production workflow.

[Artificial Analysis independently lists](https://artificialanalysis.ai/models/mistral-large-4) the high-reasoning version at 26 on its Intelligence Index v5.0. That broader measure combines nine evaluations. It is a different metric from the Coding Agent Index, so the two headline numbers cannot be compared directly or averaged into a meaningful overall score. The featured scorecard keeps their sources and meanings separate.

For a prospective user, the sensible next step is a task-specific comparison: the same repository, tools, acceptance criteria and spending limit across candidate models. A system that produces a plausible patch still needs tests and human review. PromptsZone has not run the model or independently reproduced Mistral's performance claims.

## Open weights come with a commercial boundary

The [Mistral Ecosystem Open-Weight License](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) permits broad use and modification but adds a restriction for operators of model-as-a-service businesses. Those operators need a separate commercial license for commercial use. The text defines the category and lists exceptions, including internal use and access through Mistral's official products. Businesses should review the actual terms for their intended deployment rather than assume the release is an unrestricted MIT model.

## Why this release matters

Downloadable weights shift an important decision from the vendor to the operator: where inference runs, how the model is adapted and which controls surround it. They also shift responsibility for updates, access controls, monitoring and capacity planning. That trade-off can be valuable where data location or customization is a priority, even without an across-the-board benchmark lead.

The October 29 release is therefore a concrete availability milestone, not simply a repeat of the earlier preview announcement. Its value will depend on whether the new deployment freedom outweighs the operational burden for the organizations using it.

Scorecard: PromptsZone original graphic. Data: [Artificial Analysis, Intelligence Index v5.0](https://artificialanalysis.ai/models/mistral-large-4), and [Mistral's release announcement](https://mistral.ai/news/mistral-large-4-oss/), checked October 29, 2026. The coding result is reported by Mistral; no chart values were generated or estimated.