Chinese open-weight AI models have become the leading distribution ecosystem, according to Nathan Lambert's congressional-briefing expansion, with downloads and tracked open-model inference use outpacing US counterparts. The story matters because open-model competition is no longer only a benchmark contest: the model people download, fine-tune, host and teach with can become the platform everyone else builds upon.
Key facts
- Lambert says Chinese open models have about 3.2 billion downloads, roughly twice the US total.
- He says Chinese models account for more than 80% of the open-model OpenRouter usage he tracks.
- The briefing was prepared for congressional members and staff, not verified as formal hearing testimony.
- Primary source: Lambert's analysis.
Lambert's mechanism is a flywheel. A lab ships public weights frequently; developers use them because they are cheap and capable; those weights become bases for fine-tunes, inference-provider defaults, academic work and startup products; that adoption brings feedback and further makes the model family a default. He writes that Chinese-model downloads passed US models in July 2025 and estimates a current Chinese lead of about 1.6 billion. The ATOM report independently supports the direction of travel, saying Chinese models overtook US counterparts in 2025 and widened the gap.
The 80% claim is easily abused. It means more than 80% of open-model usage tracked by Lambert on OpenRouter, not all OpenRouter traffic, all enterprise inference or all global AI use. Download counts also do not equal active deployments. Think of it as measuring which open-source operating systems developers choose to install, not a census of every computer in the world. It is nonetheless a leading indicator because installation produces skills, tutorials, derivatives and switching costs.
Lambert argues that frequent releases, high demand for coding/agent tasks and open distribution explain more than simplistic claims that China merely copied US models. He estimates leading Chinese open models are around two to five months behind the closed US frontier, while leading US open models sit six to nine months back. That is analysis, not a settled fact. In a related podcast discussion, JS Denain offered a six-to-eight-month public-model gap and called shorter estimates potentially “overfitting-y.”
The policy implication is not automatically restriction. Lambert's stated recommendation is stronger domestic investment in open models, since US companies already use Chinese weights and a ban does not build an alternative developer base. The caveat is practical: origin, security review, data residency, legal constraints and support quality may make a leading download unsuitable for a given organization. The signal worth acting on is ecosystem momentum: a durable AI advantage can accumulate through distribution and derivative work even while the closed frontier remains elsewhere.
Originally published on Ground Truth, where every claim is checked against the primary source.
