Mistral is back with something enormous. Mistral Large 4, nicknamed Le Chonk, is a 1.05 trillion parameter Mixture of Experts model with 49 billion active parameters, native image input, and a 1 million token context window. The API preview is live. Open weights are promised for the end of October 2026.

The number everyone will quote is 1.05T. The number worth sitting with is 3,800.

Owning the cluster

That’s how many NVIDIA Grace Blackwell GPUs Mistral used to train this, in its own European datacenters. Not rented capacity from a hyperscaler, not a compute-for-equity arrangement with a cloud partner. Mistral’s own metal, on the continent, under EU jurisdiction.

For the past two years the European AI story has mostly been about regulation and about whether anyone over there could actually build at frontier scale without borrowing someone else’s GPUs. A 3,800-chip Blackwell cluster is a real answer to that question. It is also a capital commitment that will look either prescient or catastrophic depending on how the next 18 months of model economics shake out, and nobody gets to know which yet.

If you care about where your inference physically happens, this is the most interesting release of the quarter. If you only care about leaderboard position, you’ll probably find it unremarkable.

”Open weights” is doing some heavy lifting

Here’s where I get skeptical. 49B active parameters means the compute per token is modest, roughly in the range of a mid-sized dense model. But sparse activation does nothing for memory. You still have to hold all 1.05 trillion parameters somewhere fast.

Do the arithmetic. At 8-bit that’s about a terabyte of weights before you account for KV cache, and the 1M token context window means the KV cache is not a rounding error. Quantize hard to 4-bit and you’re still looking at half a terabyte. That is multiple 8-GPU nodes, minimum, for one model instance.

So when the weights drop at the end of October, who downloads them? Cloud providers, a handful of well-funded startups, some national labs, and maybe three people on Reddit with a decommissioned DGX rack. The local-LLM crowd that made Mistral 7B a phenomenon is not running this on a Mac Studio.

That doesn’t make the release worthless. Open weights at this scale mean you can audit the thing, fine-tune it on data you’d never send to an API, and host it in a jurisdiction you pick. Those are real, and they’re the reasons sovereignty-conscious buyers will care. Just don’t confuse “open” with “accessible.”

Simon Willison, who got hands on the preview, flagged that the reasoning controls are limited to two modes. Thin, compared to the per-request reasoning budgets competitors now expose. Worth watching whether that changes before the weights ship.

Two things I don’t have answers to yet: which license the weights land under, and whether the released checkpoint matches the model currently serving the preview API. Both get settled in roughly three weeks.