# Mistral Large 4 on Hugging Face: Open Weights, License and What You Need to Run It

URL: https://whatisagenticai.net/blog/mistral-large-4-open-weights/
Published: 2026-10-29
Publisher: Little Tough LLC, maker of AI Agent Assistant: Agentic AI (https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260)

Mistral Large 4 is now on Hugging Face as open weights (Oct 29, 2026). The license, the new lower API price, the GPUs you need, and is it really open source.

Mistral Large 4 is now on Hugging Face. On October 29, 2026, Mistral published the open weights of its 1-trillion-parameter model, the one everybody calls “Le Chonk,” three weeks after it first showed up as a paid preview.

That raises three practical questions. Is Mistral Large 4 open source now? Can you download it and run it yourself? And did anything change from the preview? Here are the answers, with the numbers that matter.

**TL;DR:** Mistral Large 4’s weights went live on Hugging Face on October 29, 2026, under the Mistral Ecosystem Open-Weight license, a modified MIT license that is free for personal and commercial use except for companies selling it as a hosted model service. The released version had three more weeks of training, so it scores higher than the preview. The API price also fell to $0.70 per million input tokens and $2.10 per million output tokens. Running it yourself takes a server with several data center GPUs, not a laptop.

## Is Mistral Large 4 on Hugging Face?

Yes. The model lives at mistralai/Mistral-Large-4-1T-A52B on Hugging Face. The name tells you the size: about 1 trillion total parameters, with about 52 billion active for each token (49 billion if you leave out the embedding and output layers). Mistral’s release post is titled “Le Chonk: Built for Coders, Trusted by Defenders,” and The Next Web confirmed the weights went up on October 29 at 2:00 pm UTC.

One checkpoint covers both normal answers and step-by-step reasoning. You pick a reasoning level per request: low, medium or high (high is the default). The “low” level replaces the old “none” setting from the preview.

New to the model? Start with our explainer, [what is Mistral Large 4](https://whatisagenticai.net/blog/what-is-mistral-large-4/), which covers the nickname, the architecture and the preview launch.

## Is Mistral Large 4 open source?

It is open-weight, with a license close to open source. Mistral released it under the **Mistral Ecosystem Open-Weight license**, which the model card describes as a modified MIT license. In plain words:

- You can use, change, fine-tune and sell products built on it, for free.

- The one catch: a business that runs it as a “Model-as-a-Service” offering, meaning it sells access to the model itself, needs a separate commercial license from Mistral.

That is less open than Mistral Large 3, which used plain Apache 2.0. It is still far more open than closed models like GPT-6 Astra or Claude Opus 5.5, which you can only rent through an API. Mistral has not released its training data, so by the strict definition used by open source groups, this is “open weights,” not fully open source.

## What changed from the preview?

The public weights are not the same model as the October 6 preview. Mistral says it ran about three more weeks of reinforcement learning, mostly on coding and cybersecurity. All numbers below are Mistral’s own unless noted.

| Test | Preview (Oct 6) | Open weights (Oct 29) |
| --- | --- | --- |
| Terminal-Bench 4.0 (coding in a terminal) | 28% | 41% |
| B3 agent security (attacks resisted) | 93.3% | 97.1% |
| FinWorkBench (finance tasks) | 67.4% | 73.3% |
| API input price (per million tokens) | $1.36 | $0.70 |
| API output price (per million tokens) | $4.18 | $2.10 |
| Cached input (per million tokens) | $0.14 | $0.07 |
| License | Not announced | Modified MIT (Mistral Ecosystem Open-Weight) |

Mistral also reports 48% on the Artificial Analysis Coding Agent Index, roughly tied with China’s GLM-5.3 while using 47% fewer tokens per task, and a 96.6% refusal rate on malicious hacking requests. On The Next Web’s chart of Terminal-Bench 4.0, it sits just ahead of GLM-5.3 (40%), with DeepSeek V4.1 Flash at 31% and Kimi K3 at 21%.

## Mistral Large 4 hardware requirements

This is where “free to download” meets reality. The full model is huge, so it needs a multi-GPU server:

- **Full checkpoint (FP8):** Mistral’s own vLLM example on the model card splits the model across 8 GPUs. AlphaSignal notes it needs “substantial aggregate GPU memory, storage, and interconnect bandwidth.”

- **Smaller NVFP4 version:** a compressed checkpoint aimed at Nvidia Blackwell chips. The Next Web reports this version runs on four Nvidia B200 or B300 GPUs.

- **Software:** vLLM is recommended. SGLang, Hugging Face Transformers and llama.cpp (GGUF files) are also supported.

Four B200s is still a data center machine, so for a home computer or a Mac, Mistral Large 4 is out of reach today. Mistral told The Next Web that smaller distilled versions, including one that runs on a laptop, are coming “in the coming months.” No date has been given.

## Where can you use Mistral Large 4 without a server?

Per Mistral’s release post, hosted versions are available on:

| Where | What it is |
| --- | --- |
| Mistral Studio and the Mistral API | Mistral’s own paid API and playground |
| OpenRouter | One API key for many models |
| Microsoft Foundry | Microsoft’s cloud AI platform |
| Nvidia | Hosted endpoints and an NIM container |
| Vibe (chat.mistral.ai) | Mistral’s coding agent |

A special cybersecurity version with fewer limits is not public. Mistral is giving it to vetted enterprise partners, including Orange Cyberdefense, for tasks like penetration testing on systems they own or are allowed to test.

## Is it the best open model now?

Depends on who you ask. Mistral says it leads open-weight models on its coding and cyber tests and that human reviewers at Surge preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Independent rankings are more mixed. The Decoder’s headline today says it is a strong coding model that “still trails China overall,” and earlier Artificial Analysis results placed several Chinese open models ahead of it on general intelligence. What most sources agree on: it is the strongest open-weight model made outside China. For how the top closed models compare, see our [Gemini 4 Carbon vs Claude Opus 5.5](https://whatisagenticai.net/blog/gemini-4-carbon-vs-claude-opus-5-5/) breakdown.

## What this means for you

For banks, governments and security teams in Europe, this is a big deal. They can now run a top-tier model on their own machines, read every byte of it, and never worry about an outside company switching it off. That is exactly what Mistral is selling.

For one person with an iPhone, the release changes little. You will not run a 1-trillion-parameter model at home, and you probably do not want to manage API keys either. What you want is the work done: a trip planned, a report drafted, a topic researched. That is the job of [agentic AI](https://whatisagenticai.net/what-is-agentic-ai/). **Agentic AI** puts a team of up to four expert agents on your iPhone and iPad, with no account needed, so you get the results without picking or hosting a model.

## Quick answers

**Is Mistral Large 4 open source?** It is open-weight under the Mistral Ecosystem Open-Weight license, a modified MIT license. It is free to use and sell products with, but companies that sell it as a hosted model service need a separate license. Training data is not released.

**Is Mistral Large 4 on Hugging Face?** Yes, since October 29, 2026, at mistralai/Mistral-Large-4-1T-A52B, plus a smaller NVFP4 checkpoint for Nvidia Blackwell GPUs.

**What are the Mistral Large 4 hardware requirements?** Mistral’s example uses 8 GPUs for the full model. The NVFP4 version runs on four Nvidia B200 or B300 GPUs, per The Next Web. It does not fit on a laptop or a single consumer GPU.

**Is Mistral Large 4 free?** The download is free. Using Mistral’s API costs $0.70 per million input tokens and $2.10 per million output tokens, about half the preview price.

**Can I run Mistral Large 4 with Ollama or LM Studio?** GGUF builds for llama.cpp are linked from the model card, but the model is still about a trillion parameters, so you need server-class memory. Smaller distilled versions are promised but not dated.

Words like “open weights,” “parameters” and “distillation” are explained in our [agentic AI glossary](https://whatisagenticai.net/agentic-ai-glossary/).

**Sources:** [Mistral: Le Chonk, Built for Coders, Trusted by Defenders](https://mistral.ai/news/mistral-large-4-oss/) · [Hugging Face model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B) · [The Next Web](https://thenextweb.com/news/mistral-large-4-open-weights-release) · [AlphaSignal](https://alphasignal.ai/news/mistral-drops-large-4-a-1-trillion-parameter-open-coding-agent) · [Mistral: Introducing Mistral Large 4](https://mistral.ai/news/mistral-large-4) · [DeepLearning.AI The Batch](https://www.deeplearning.ai/the-batch/mistrals-skills-match-its-global-appeal/)