# Mistral Large 4 packs a trillion parameters and ships as open weights this month

> Mistral Large 4 is a 1.05 trillion parameter mixture of experts model in public preview, with downloadable weights promised by the end of October 2026.

- Published: 2026-10-12
- Topic: [AI](https://futuretechnologyhq.com/topic/ai/)
- URL: https://futuretechnologyhq.com/article/mistral-large-4-preview-1-trillion-parameters-open-weights/
- Publisher: [Future Technology](https://futuretechnologyhq.com/), free to read. Quote as "Future Technology" with a link to the article.

## Key points

- Mistral Large 4 entered public preview on 6 October 2026 as a multimodal mixture of experts model with 1.05 trillion total parameters.
- Roughly 49 to 52 billion parameters are active per token, so the model runs cheaper than its headline size suggests.
- Access today is through a hosted API on Mistral Studio; downloadable weights are promised by the end of October.
- Reinforcement learning is still underway and every benchmark figure so far comes from Mistral.

## Article

1.05 trillion parameters, and the plan is to let you download them. Mistral opened a public preview of Large 4 on 6 October, and the company says the weights will follow by the end of the month.

Mistral calls it the largest and most capable model it has built, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres. Co-founder Guillaume Lample called it "by far the strongest open-weight model from the US or Europe".

## What Mistral Large 4 is

Large 4 is a multimodal mixture of experts model. The trillion figure is the total parameter count, but only a slice of the network fires for any given token. Early reports put that active count at 49 billion. BetaNews notes the official documentation now says about 52 billion, so treat it as roughly 50 billion until Mistral's own pages settle.

That split is the whole trick. A mixture of experts model stores a very large amount of knowledge, then routes each token through a small subset of it, so inference costs look closer to a 50 billion parameter model than a trillion parameter one. Our explainer on [open weight versus closed AI models](/article/open-weight-vs-closed-ai-models-explained/) covers why that architecture has become the default for anyone releasing weights.

Mistral's documentation lists a one million token context window. For now, access is through the preview API on Mistral Studio, where the documentation lists $1.36 per million input tokens and $4.18 per million output, halved to $0.68 and $2.09 during a launch sale.

## Open weights, with an asterisk for now

Mistral calls Large 4 an open weight model, but the public preview is the hosted version only, so for now it is open weight in name rather than as a file. The gap between "open weight" as a promise and open weight as a file you can download is worth remembering when the release date slips, if it does.

Mistral has also said reinforcement learning is still underway. In plain terms, the model you are testing in October is not the finished model, and every benchmark number released so far is Mistral's own, with the preliminary label attached. Independent evaluations will come after the weights do.

## Can you run it at home

Not this one, or at least not in full. A trillion total parameters at 8-bit precision is about a terabyte of weights, and even aggressive 4-bit quantisation leaves you around 500 GB. That is multi-GPU server territory, not a desk. Our [local LLM VRAM requirements guide](/article/local-llm-vram-tier-list-2026/) walks through the maths, and the piece on [unified memory and why 128 GB matters](/article/unified-memory-explained-128gb-local-ai/) explains why Macs with big memory pools have become the hobbyist workaround for mid-sized models.

Where the open release matters is one step up from the hobbyist. Companies and research groups that can rent a few nodes get a frontier scale model they can fine-tune, inspect and host on their own terms, which is the audience Mistral has always been chasing.

## What to watch

The end of October deadline, and what arrives with it. If the weights land with a permissive licence and a clear parameter sheet, Large 4 becomes the reference open model for the rest of the year. If the date slips or the licence restricts commercial use, the "open weight" claim will get a harder look. The second thing is whether the active parameter number settles at 49 or 52 billion, because that decides what it costs to serve.

## Sources

- [Mistral Large 4 preview and open weights plan (BetaNews)](https://betanews.com/article/mistral-large-4-preview-open-weights/)
- [Mistral Large 4 debuts as 1T parameter open weight model (VKTR)](https://www.vktr.com/ai-platforms/mistral-large-4-debuts-as-1t-parameter-open-weight-model/)
- [Mistral opens public preview of Large 4 (Let's Data Science)](https://letsdatascience.com/news/mistral-debuts-large-4-open-weight-model-preview-cbf9e0b5)