Chinese Open-Weight Models Now Lead US Rivals by 2-6 Months, Congressional Briefing Shows

The briefing

Nathan Lambert, a researcher at the Allen Institute for AI (Ai2) who helped build the Olmo model series, briefed Congressional members and staff on the state of open-weight AI models in the context of U.S.-China competition. He has published his prepared remarks publicly via Interconnects.

The core claim: Chinese open-weight models have decisively overtaken American open-weight models on both adoption and capability metrics, and the gap is not closing.

The numbers

According to Lambert, China took the lead in Hugging Face downloads in July 2025, driven largely by Alibaba's Qwen models. Since he launched the American Truly Open Models (ATOM) Project tracker in August 2025, China's download lead has grown to approximately 1.6 billion, out of a total 3.2 billion downloads — meaning Chinese models now account for roughly twice the downloads of American open-weight models.

On the Artificial Analysis Intelligence Index (AAII), as of September 14, 2026, the top three models are all Chinese: Z.ai's GLM-5.3 (45), Moonshot AI's Kimi K3 (44), and GLM-5.3-Flash (42). The leading American open-weight models — Thinking Machines' Inkling and Inkling Small, and Nvidia's Nemotron 3 Ultra — score 26, 26, and 23 respectively. Lambert states the top American open models trail 15 Chinese-made models on this index.

Lambert estimates Chinese open-weight models are roughly 2-5 months behind the closed American frontier (OpenAI, Anthropic), while American open-weight models trail 6-9 months behind that same closed frontier. Chinese labs are closest to parity in agentic coding, where commercial demand is highest, and furthest behind in open-ended scientific domains like physics and biology.

Why the gap exists

According to Lambert, faster release cadence is a major factor: because all labs are improving continuously, a model released later captures more accumulated progress, mechanically inflating benchmark scores for labs that ship faster — a category currently dominated by Chinese firms including Z.ai and Moonshot AI.

He also points to distillation from closed American APIs as a contributing but limited factor. Lambert estimates that even if distillation were fully blocked — for instance through know-your-customer verification at OpenAI and Anthropic — the resulting gap increase would only be 1-2 months, arguing distillation is not the primary driver of Chinese progress.

A newer trend Lambert highlights: Chinese labs, including Moonshot AI and Z.ai, shifted in 2026 from building agentic RL training data in-house toward purchasing challenging RL environments from both American vendors and new Chinese startups.

Policy tension

Lambert raises a specific case tied to security policy debates: Hugging Face reportedly used a Chinese open-weight model to analyze a cyberattack because closed American models declined to respond to the relevant queries. He uses this to argue that restricting access to capable Chinese open-weight models would primarily harm American businesses that already depend on them, not reduce risk, since open software is structurally difficult to gatekeep once released.

What this means

This is one researcher's testimony and self-maintained tracking data (ATOM Project, AAII), not independently audited figures — treat the specific benchmark scores and download totals as claims from Lambert's analysis rather than verified consensus numbers. That said, the directional story — Chinese labs like Z.ai, Moonshot AI, and Alibaba's Qwen team leading open-weight releases since mid-2025 — is consistent with broader industry observation.

The practical implication for developers: if you're building on open-weight models today, the strongest options increasingly come from Chinese labs, not American ones. That creates a real policy dilemma Lambert doesn't resolve — restricting Chinese open models to manage risk could just push American developers toward dependency on Chinese infrastructure with no corresponding security gain, since the models are already downloadable. Watch whether US labs like Nvidia, Thinking Machines, Arcee AI, Poolside, and IBM close the gap in coming release cycles, or whether the 2-6 month Chinese lead becomes structural.