Garry Tan Urges US Open-Weight AI Labs to Distill Frontier Models to Counter China

TL;DR

• Y Combinator CEO Garry Tan says small U.S. open-weight labs should openly distill frontier models from American giants like OpenAI, Anthropic, Meta and xAI, just as Chinese labs did to catch up.
• Tan argues the tactic is the fastest way to break U.S. startups' growing dependence on Chinese open models like DeepSeek and Qwen and build a homegrown alternative.
• The proposal reignites debate over model terms of service, copyright, and whether Washington should bless distillation as a national competitiveness strategy.

Borrowing Beijing's Playbook

Garry Tan wants American AI startups to fight fire with fire.

The Y Combinator CEO said this week that smaller U.S. open-weight labs should be encouraged — not punished — for distilling frontier models built by U.S. giants, mirroring the exact tactics Chinese labs used to vault to the front of the open-model race.

In posts on X and in interviews around Y Combinator's latest batch, Tan argued that distillation, where a smaller student model is trained on the outputs of a larger, more capable teacher model, is how DeepSeek, Alibaba's Qwen team, Moonshot AI and others closed a multi-year gap with Silicon Valley in months. Now, he says, American open builders should do the same thing with American teachers.

Instead of U.S. startups fine-tuning Qwen or DeepSeek because they are the best-performing open weights available, Tan wants them fine-tuning and distilling Llama, GPT, Claude and Grok — and releasing the results as American open weights for everyone to build on.

Distillation Explained — And Why It's Controversial

Distillation is not new, but it has become the most contentious practice in AI.

In simple terms, a lab pays for millions of queries to a frontier model like OpenAI's GPT-5, Anthropic's Claude or Google's Gemini, then uses those question-answer pairs to train a far smaller, cheaper model that mimics much of the larger model's reasoning and style. Done well, a 7-billion or 70-billion parameter model can approach frontier performance at a fraction of the training cost.

That is precisely what OpenAI accused DeepSeek of doing last year after the breakout of its R1 reasoning model — alleging DeepSeek improperly harvested OpenAI outputs to train its own competitor. David Sacks, the White House AI and crypto czar, at the time called it evidence of IP theft.

Tan is now flipping that framing. If distillation powered China's rise, he argues, banning U.S. startups from doing it only locks in American disadvantage. He has called on frontier labs to explicitly allow smaller U.S. labs and researchers to distill from their APIs for open-weight releases, rather than banning the practice in their terms of service.

America's Quiet Dependence on Chinese Models

At the heart of Tan's push is an uncomfortable fact for Silicon Valley: the U.S. open-weight ecosystem is losing to China.

While closed frontier leadership remains American — with OpenAI, Anthropic, Google DeepMind and xAI trading the top spots on benchmarks — the downloadable, modifiable models that power thousands of startups, agents, and on-device apps are increasingly Chinese. Qwen, DeepSeek, GLM and Kimi have dominated Hugging Face download charts through 2025 and 2026, praised for strong performance, permissive licenses and rapid iteration.

Tan and other YC partners say they see it firsthand. A growing share of YC AI startups default to Qwen or DeepSeek derivatives because there is no equally capable, truly open American alternative since Meta slowed its Llama program. That, Tan warns, creates economic and national security risk: sensitive U.S. products built on models subject to Chinese censorship, Chinese licensing shifts, and potential Beijing leverage.

A homegrown open-weight stack distilled from U.S. frontier models, he argues, would give American founders a patriotic, high-performance default.

A Call For a Homegrown Open-Weight Ecosystem

Tan's vision is not a single model, but a flywheel.

Frontier giants provide the teacher brains. Small labs — the likes of Nous Research, EleutherAI, Allen Institute for AI, Together AI ecosystem players, and YC-backed upstarts — distill them into lean, specialized, open-weight models for coding, science, voice and robotics. Those models then get remixed by thousands of startups, creating the robust, decentralized ecosystem that made early Llama so powerful.

To get there, Tan suggests three shifts: frontier labs should create a formal distillation carve-out for U.S.-based open research, policymakers should treat open weights as critical infrastructure rather than a safety threat, and Big Tech cloud credits should flow to distillers, not just to training new foundation models from scratch.

Critics note the irony: asking OpenAI and Anthropic to subsidize their own open-source competitors is a tough sell, and wholesale distillation raises unresolved questions about copyright, creator consent and model collapse.

What Comes Next

Whether Tan's idea becomes policy or remains provocation, it signals a major shift in Silicon Valley thinking about China.

For two years, the dominant response to DeepSeek was to condemn distillation as cheating. Tan's argument — that America should legalize and weaponize it at home — reframes it as industrial strategy.

With Washington debating export controls, open-model restrictions and the upcoming AI action plan implementation, Tan's proposal puts Y Combinator squarely on the side of permissive, pro-distillation, pro-open-weights policy. His message to founders is blunt: stop waiting for a perfect American Llama replacement, start distilling American frontier models today, and out-China China.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.