Congressional briefing lays out China's lead in open-weight AI models

A briefing prepared for Congress

The AI newsletter Interconnects has published the prepared remarks from a briefing given to members of Congress and their staff on open-weight AI models, framed around competition between the United States and China. The author spent two and a half years at the Allen Institute for AI helping build its OLMo models and presents the remarks as an overview of where openly available language models currently stand.

Open-weight is not the same as open-source

The briefing begins with definitions that policy debates often blur. Open-weight models publish their trained parameters for inspection and downstream use under licenses; Meta's Llama, Alibaba's Qwen, Google's Gemma and DeepSeek's releases are the familiar examples. Genuinely open-source models go further, shipping everything required to reproduce a model, including training code and training data. According to Interconnects, the most prominent fully open-source efforts — AI2's OLMo, OpenAthena's Marin and EleutherAI's Pythia — all come from American non-profits.

The categories form a spectrum rather than separate buckets. Nvidia's Nemotron models, for example, publish substantial training data under permissive licenses, making them more open than most open-weight releases without qualifying as fully open-source. Closed API-only models likewise vary in how much they reveal.

The numbers behind China's lead

Per the briefing, Chinese companies have led the open-weight field since roughly April 2025, overtaking the early American advantage built on Meta's Llama. On Hugging Face downloads, China moved ahead in July 2025, largely on the strength of Alibaba's Qwen family. The author tracks this with the American Truly Open Models (ATOM) project, first published in August 2025; since then China's download lead has grown to roughly 1.6 billion, with about 3.2 billion Chinese model downloads in total, twice the American figure.

Benchmarks point the same way. On the Artificial Analysis Intelligence Index, as of September 14, 2026, the top three open models are Chinese: Z.ai's GLM-5.3 and GLM-5.3-Flash, scoring 45 and 42, and Moonshot AI's Kimi K3 at 44. The leading American entries — Thinking Machines' Inkling and Inkling Small at 26 each, and Nvidia's Nemotron 3 Ultra at 23 — sit behind fifteen Chinese models on that index. The top US models were released in June and July 2026 and are updated less often; Chinese labs reached comparable scores two to six months earlier, with releases such as GLM-5 and DeepSeek V4 Pro. Newer American participants including Arcee AI, Poolside and IBM are not closing the distance quickly.

How far everyone is from the frontier

Interconnects estimates Chinese open-weight models sit about two to five months behind the closed American frontier held by OpenAI and Anthropic, while American open-weight models are six to nine months back. Chinese releases are strongest where user demand is clearest, notably agentic coding, and weaker on open-ended scientific problems in fields like physics and biology. GLM-5.2 and Kimi K3 are described as having reached an agentic-capability level comparable to the one Claude Code crossed in December 2025.

Part of the Chinese lead is attributed to release cadence: since every lab improves continuously, a model finished later captures a stronger performance snapshot, and a somewhat narrower task focus makes benchmark results look better than they are. Distillation is downplayed as an explanation — the briefing estimates that fully blocking it, for instance through know-your-customer checks at OpenAI and Anthropic, would widen the gap by only a month or two. The document also records a 2026 shift among Chinese labs from building training data in-house to buying cutting-edge data, such as reinforcement learning environments for agentic tasks, from established American companies and new Chinese startups.

What it means for policy

The briefing warns that restricting Chinese open-weight models would mostly hurt American users. During the recently documented OpenAI–Hugging Face cybersecurity incident, Hugging Face analysed the attack using a Chinese open-weight model because closed models declined to answer its queries. Cutting off access to such models would therefore set back American businesses, the argument runs, and managing the risks of open weights comes down largely to preparing the ecosystem. The recommended course is continued US investment in open models, both to coordinate on risks that are global in nature and to avoid a lopsided dependence on Chinese-built weights.

Why it matters

This is a practitioner's assessment entering the Congressional record while US AI policy is still being shaped. It reframes the debate from containing open models to preparing for them: if the numbers hold, American businesses are increasingly depending on Chinese-built weights, and restrictions would tax domestic users rather than foreign labs. It also gives legislators concrete evidence — download figures, benchmark scores and frontier-gap estimates — for decisions about funding open research and regulating model releases.