Two things are true at the same time right now. The most capable AI models in the world are being built in American labs. And when I test open-weight models for this site — the ones you can actually download, fine-tune, and run yourself — the strongest options increasingly come from somewhere else. That gap is exactly what Garry Tan is pointing at, and I think he’s onto something worth taking seriously.
Tan, who runs Y Combinator, is calling on smaller American open-weight AI labs to distill frontier models. The goal, as he frames it, is to build a strong alternative to Chinese AI dominance in the open-weight space, and he sees this as crucial for advancing American AI research by 2026.
As someone who spends most of my week benchmarking AI tools and telling you which ones deserve your money and which ones don’t, I want to unpack why this matters for people who actually build things — not just for people who write policy memos.
What Distillation Actually Means for Builders
Quick refresher for anyone new to the term. Distillation is the process of training a smaller model to imitate a larger, more capable one. The big model acts as a teacher, the small model as a student. You lose some capability, but you gain something enormously practical — a model that’s cheap to run, fast to serve, and possible to deploy without a data center budget.
From a toolkit reviewer’s perspective, distilled models are where the real action is. Frontier models are impressive in demos, but most working developers I talk to don’t run frontier models for everything. They run smaller models for the boring, high-volume tasks — classification, extraction, summarization, routing. The economics demand it.
So when Tan asks American open-weight labs to distill frontier models, he’s not asking for a science project. He’s asking for the exact category of tool that developers reach for every single day.
The Honest Assessment
Let me put on my reviewer hat and evaluate this idea the way I’d evaluate a product.
What works about the pitch
The logic is sound. If the best teacher models are American, then American labs are well positioned to produce strong student models. Distillation is also one of the few strategies available to smaller labs that can’t afford frontier-scale training runs. It’s a way for a lean team to punch above its weight — which, not coincidentally, is the kind of company Y Combinator exists to fund.
There’s also a real user need here. Builders want open weights. They want models they can inspect, modify, self-host, and ship without usage restrictions dictated by an API provider. Right now, if you want that combination of openness and quality, your options are limited, and the strongest ones often aren’t American. Tan’s push addresses a genuine hole in the market.
What I’d flag as a reviewer
Distillation isn’t free. Someone has to decide what the frontier labs think about their models being used as teachers, and the terms of service around this vary. There’s also the question of differentiation. If a dozen small labs all distill from the same frontier models, do we get a dozen meaningfully different tools, or a dozen slightly different flavors of the same student? Sameness is already a problem in this market — I review these tools, and believe me, a lot of them blur together.
And the 2026 timeline Tan attaches to this is ambitious. Training pipelines, evaluation, safety work, and actual release engineering take time, even when you’re standing on a teacher model’s shoulders.
Why I’m Cautiously in Favor
Here’s my honest take. The open-weight space needs more serious American entrants, not for flag-waving reasons but for practical ones. Competition makes tools better. Every time I’ve seen a category of AI tooling get a strong new competitor, quality went up and prices came down for everyone. More labs distilling, releasing, and iterating means more options landing on my review bench — and more genuinely useful choices landing in your stack.
The alternative to Tan’s proposal isn’t neutrality. It’s a future where the open-weight tier — the tier most builders actually live in — is shaped primarily by labs outside the US. Whether that bothers you politically or not, it should bother you practically if you want a diverse, competitive market of tools.
So my verdict, in typical review format: the idea is solid, the need is real, the timeline is aggressive, and execution is everything. If American open-weight labs answer Tan’s call, I’ll be here testing every release — and
🕒 Published: