Mistral Large 4 Landed October 6. The Weights Didn't.

Mistral Large 4 hit public preview on October 6, 2026, and you still can't download it.

The model that Mistral's own announcement calls "Unofficially ML4, very officially: le Chonk" is a 1-trillion-parameter open-weight, natively multimodal model with 52 billion active parameters, a 1.6-billion-parameter vision encoder. And a 1-million-token context window. Today the only way to run it is the preview API on Mistral Studio. Mistral says the downloadable weights arrive by the end of October 2026, packaged with additional architecture details, extra benchmarks, and its post-training methodology. That gap between "announced open" and "actually downloadable" is the whole story. And it should shape what you do with your stack this month.

The Specs: 1.05 Trillion Total, 52 Billion Active

The announcement rounds up and calls ML4 "a 1 trillion-parameter natively multimodal model with 52 billion active parameters." The model documentation is more precise: 1.05 trillion total parameters, 52 billion active, plus a separate 1.6 billion-parameter vision encoder, all built on a granular Mixture-of-Experts design. The changelog confirms the 1M context window on top.

The two numbers that matter to you are total and active. With a MoE architecture, you store and move all 1.05 trillion parameters. But only 52 billion fire on any given token. That split is why frontier MoE models feel faster and cheaper per token than their raw size suggests. And it's also why "open-weight" here means a serious storage and memory commitment even if the per-token math looks friendly. The vision encoder tells you the model reads images natively rather than through a bolted-on pipeline, which matters if your automation touches screenshots, documents, or UI elements.

I'm deliberately not translating any of that into a hardware bill. Nobody has run the real thing locally yet because the weights aren't out. And a spec sheet without a checkpoint is a paper exercise. When the download lands, the first honest numbers will come from people with actual GPUs, not from launch-day blog posts.

Every Claim So Far Comes From Mistral

Here's my contrarian read: the strongest claims about le Chonk currently have exactly one source. And it's the enterprise selling it. The announcement states that "on critical enterprise workloads, including cybersecurity, finance and law, we find it to be state-of-the-art among open models." The coding story is two named evaluations, DeepSWE and Terminal Bench 4.0, listed in the launch post without scores attached in anything I can verify.

That's not an accusation, it's a stage of a rollout. Mistral has explicitly promised more: additional benchmarks and post-training methodology arrive alongside the weights. So the defensible position right now is "promising, self-reported, unverified," and the indefensible position is rebuilding anything around it.

My rule after enough of these launch cycles is blunt: nothing graduates from "interesting" to "infrastructure" until third parties have pulled the weights and run their own workloads against them.

There's as well a quieter signal in the specs worth respecting. A general-purpose multimodal model with a 1M context window is aimed at exactly the workloads small teams care about, long documents, large codebases, multi-step pipelines.

But a listed context window and a usable one are different things. And I've watched enough launches to know the usable number shows up in practice, not in docs.

Vetted Testers Get a Looser Model Than You

The most under-discussed part of this launch is the access structure. Until the weights drop, Mistral says it is "red-teaming the model in real-world settings with cybersecurity leaders, vetted partners.

And state authorities, who will access the same model with reduced moderation and expanded cyber capabilities."

Read that as a two-tier system.

Selected partners and state authorities test a version with reduced moderation and expanded cyber capabilities, while everyone else gets the standard preview API.

And the downloadable weights come after.

This is the emerging template for "open" launches: preview first, controlled red-teaming second, weights last. It's a reasonable safety posture. And it's too a reminder that "open-weight" is a delivery promise, not a description of what exists today. The legal center entry stamps the October 6, 2026 release date. And the changelog still says "Open weights are coming soon."

When the download link goes live, three things deserve your attention before any enthusiasm: the license terms, the post-training methodology Mistral promised, and the additional benchmarks.

Open weights with a restrictive license is a very other product than open weights you can fine-tune and ship.

What This Changes for a Lean Operation

If you build and ship AI automation, downloadable frontier weights change your build-versus-API math in one specific way: the exit option becomes real. API-locked models mean a pricing change or a policy change on someone else's calendar becomes your emergency. Weights on hardware you control mean your data stays put and your fallback exists even if the vendor's roadmap turns.

That's the entire sovereignty argument, and it's why the end of October matters more than any benchmark headline.

For a solo operator or small team, though, temper that with the storage reality of a 1.05 trillion-parameter model and the fact that nobody has published what it actually costs to serve this thing.

My honest recommendation is a waiting strategy with a work component: go try the preview API on Mistral Studio now, since your real prompts will tell you more than any leaderboard.

And keep your production stack on whatever is already earning its keep. Write down what the preview does well and where it fails. So when the weights land you're comparing against your own baseline instead of a vendor's.

The countdown is short and verifiable. Mistral has committed to weights by the end of October 2026. And that's a date you can hold them to on a public calendar.

Here's your move: this week, run your three ugliest production prompts against the Mistral Large 4 preview on Mistral Studio and note exactly where it breaks.

Then set a reminder for the end of October.

And when the weights drop, read the license before anything touches your roadmap.