Mistral Opens Preview of Trillion-Parameter Large 4, Promises Open Weights by Month's End

Mistral AI went public with Large 4 on October 6, skipping the usual slow trickle of benchmark screenshots by dropping direct API access the same day as the announcement. The architecture decision worth naming: it's a mixture-of-experts model, which means the trillion-parameter headline is real but the compute bill per request is not what that number implies.

<cite index="27-16,27-17">Mistral Large 4 uses a mixture-of-experts design, meaning the full model contains a large pool of parameters while each request activates only part of that pool. Mistral lists 1.05 trillion total parameters and 49 billion active parameters.</cite> That active-parameter figure is what actually hits the GPU on a given inference call. It's a meaningful distinction that the company's own marketing elides when it rounds to "1 trillion."

<cite index="25-8">The model is natively multimodal and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral's own European datacenters.</cite> That infrastructure detail matters geopolitically: Mistral is leaning into European digital sovereignty as a selling point, and training on company-owned compute in Europe is the proof point behind that claim.

<cite index="27-4,27-5">Mistral Large 4 is a public-preview multimodal model with a 1 million-token context window, and Mistral offers it through its moderated API while planned open weights remain under safety testing.</cite> According to the Betanews review of the launch, <cite index="19-4">downloadable weights will follow by the end of October.</cite>

The open-weights commitment is the part of this announcement that actually matters for enterprise and developer adoption, and it's also the part with the most uncertainty. Mistral hasn't said what the open weights license looks like. A model this size, running locally, requires hardware that most teams don't have on hand, so the practical audience for self-hosted Large 4 is narrower than the announcement implies. Still, the API preview gives teams a real way to evaluate the model now, which puts it ahead of most launches in this size class.

<cite index="23-7">API access is priced at $1.36 per million input tokens and $4.18 per million output tokens at preview.</cite> That's well above Mistral's smaller tiers and above where most production workloads close their cost models. Preview pricing tends to drop, but developers evaluating it for budget-sensitive applications should do the math before committing architecture decisions to it.

<cite index="20-3">The company said the model demonstrates performance across coding, agentic workflows, and multimodal understanding, and targets critical enterprise applications including cybersecurity, finance, and law.</cite> Those are the right verticals for a model that can read images and process a million tokens of context. Whether it actually outperforms the competition on production tasks in those domains is something third-party evals will have to settle, and those aren't out yet.

What's clear from the AI News report covering the October 6 launch is that Mistral's pitch is two-pronged: frontier performance and European provenance. <cite index="26-3">Mistral is a primary beneficiary of a push by the European Union and its member governments toward digital sovereignty in response to the AI arms race between the US and China.</cite> Large 4 trained end-to-end in Europe is the product form of that positioning.

The open weights, when they arrive, will tell the real story. A trillion-parameter model that anyone can download and run puts serious capability in the hands of teams that don't want to route sensitive data through any vendor's API, including Mistral's. That's either a significant shift in what's accessible or a model too large for most organizations to run without purpose-built infrastructure. Probably both, depending on who's asking.

Sources cited:
- Betanews (https://betanews.com/article/mistral-large-4-preview-open-weights/)
- Online Tech Tips (https://www.online-tech-tips.com/mistral-large-4-preview-open-weights-news/)
- Tech Journal (https://techjournal.org/mistral-large-four-preview)
- Local AI Zone (https://local-ai-zone.github.io/blog/mistral-large-4-deep-dive.html)
- Tech Insider (https://tech-insider.org/mistral-large-4-le-chonk-1-05t-parameters-2026/)
- ANI News (https://www.aninews.in/news/business/mistral-launches-1-trillion-parameter-large-4-model-plans-open-weight-release-by-month-end20261007084559/)
- AI News (https://www.artificialintelligence-news.com/news/mistral-ai-launches-large-4-preview-ahead-open-weight-release/)
- Wikipedia / Mistral AI (https://en.wikipedia.org/wiki/Mistral_AI)