Glonce

China's open-weight models now lead US rivals, briefing tells Congress

modelsThe Glonce editorsHacker News🔥70Glonce rating70Impact70Novelty65Relevance90Surprise45Players60Credibility85
China's open-weight models now lead US rivals, briefing tells Congress

The author of the AI newsletter interconnects.ai (no name is given in the piece itself, only first-person references to being invited) writes up the remarks he prepared to brief a group of U.S. Congressional members and staff on the state of open-weight AI models in the context of U.S.-China competition. The piece opens by defining terms: closed models are reached only through an API or a product, like GPT-4, Claude Opus 4.5, ChatGPT or Claude Code. Open-weight models, the far more common category, publish their weights under a license, alongside inference code, examples being Meta's Llama, Alibaba's Qwen, Google's Gemma and DeepSeek's models. True open-source models go further and release the training code and data too; the leading examples, the Allen Institute for AI's Olmo (which the author helped build over 2.5 years there), OpenAthena's Marin and EleutherAI's Pythia, are all American.

On the competitive state of play, the author says China became the clear leader in open-weight models around April 2025 and took the lead in Hugging Face downloads in July 2025, driven mainly by Qwen. Since he first published his own tracking effort, the American Truly Open Models (ATOM) Project, in August 2025, China's download lead over the U.S. has grown to about 1.6 billion, out of a 3.2 billion download total, twice America's share. On capability benchmarks the pattern repeats: as of September 14, 2026, on the Artificial Analysis Intelligence Index, Z.ai's GLM-5.3 and GLM-5.3-Flash and Moonshot AI's Kimi K3 scored 45, 42 and 44, against 26 for Thinking Machines' Inkling and Inkling Small and 23 for Nvidia's Nemotron 3 Ultra, the leading American entries, which were released only in June and July of 2026 and are updated less often than their Chinese counterparts. Models like GLM-5 or DeepSeek V4 Pro reportedly hit those same scores 2 to 6 months before U.S. labs did.

Overall, the author estimates Chinese open-weight models are about 2 to 5 months behind the closed American frontier (OpenAI, Anthropic), while American open-weight models trail that same closed frontier by about 6 to 9 months. Chinese labs are closest to parity on tasks with clear commercial demand, such as agentic coding, and further behind on open-ended scientific work like physics or biology. He attributes part of the Chinese labs' showing to faster release cycles, which flatter benchmark scores because they capture later snapshots of steady progress, plus a narrower task focus, on top of genuine strength. On distillation, he argues it matters most in new domains and does not make a universally strong model trivial to copy; he estimates that even if distillation were fully blocked, for instance through know-your-customer checks at Anthropic and OpenAI, the American-to-Chinese gap would only widen by 1 to 2 months. He also notes that Chinese labs, including Moonshot AI and Z.ai, moved during 2026 from building training-data workflows in-house to buying data, including challenging agentic reinforcement-learning environments, from both American and Chinese vendors.

On risk and policy, the piece points to growing regulatory uncertainty as open-weight capability climbs toward the frontier, citing recent frontier-model risks in areas like cybersecurity (referencing what it calls the OpenAI-Hugging Face incident, without naming the specific model involved). It notes Hugging Face itself used a Chinese open-weight model to analyze a cyberattack because closed models would not answer the request. The author's conclusion is that restricting access to the strongest Chinese open-weight models over risk concerns would mainly set back American businesses that already depend on them, so managing the risk is more a matter of ecosystem preparation than access controls, and the better path is continued U.S. investment in open models.

On adoption, OpenRouter's published usage data shows open-model weekly token throughput growing from about 1 trillion tokens in a week of September 2025 to about 80 trillion tokens per week now, with the Chinese share of that usage rising from about 70% to over 80% in the same span; the open-source coding agent OpenCode runs about 95% or more of its inference on Chinese models. The author lists prominent companies building products on Chinese open-weight models, including the legal-agent product Harvey, the coding agent Cursor, DoorDash's use of Kimi models, Airbnb's use of Qwen and Perplexity's use of DeepSeek, and says a growing number of American firms are now signing enterprise agreements with Chinese labs specifically for permission to use their models in products.

In academia, the author says Qwen is the default model for most machine-learning research, based partly on meetings with Qwen's own leadership team during a trip to China, and backs this with his own scan of every paper posted in arXiv's five most popular ML categories (cs.AI, cs.CL, cs.CV, cs.LG, stat.ML). Mentions of any open model in new papers rose from 2% in January 2023 to 50% in September 2026. In April to May 2023, about 2,600 of 12,000 new papers mentioned a prominent open model family, with about 5.5% mentioning Llama and about 1% a Chinese model; by fall 2024, Llama's peak, about 23% of papers mentioned Llama against 7.5% for Qwen. Today Llama is still mentioned in about 21% of papers, but Qwen has risen to 30%, and any Chinese open-weight model together appears in over 40% of papers against the U.S.'s 30%. The piece closes by flagging "signs of hope," including OpenAI's return to releasing open-weight models, but the captured text cuts off mid-sentence at that point, before the argument is completed.

Key facts

  • China's open-weight models (Qwen, GLM-5.3, Kimi K3, DeepSeek) hold a Hugging Face download lead over American open-weight models that has grown to about 1.6 billion since August 2025, out of a 3.2 billion total, twice America's share.
  • On the Artificial Analysis Intelligence Index as of September 14, 2026, the top three Chinese models, Z.ai's GLM-5.3 (45), Kimi K3 (44) and GLM-5.3-Flash (42), all outscored the leading American models, Thinking Machines' Inkling and Inkling Small (26 each) and Nvidia's Nemotron 3 Ultra (23).
  • Chinese open-weight models trail the closed American frontier (OpenAI, Anthropic) by an estimated 2 to 5 months; American open-weight models trail that same frontier by an estimated 6 to 9 months.
  • OpenRouter's weekly open-model token throughput rose from about 1 trillion to about 80 trillion tokens between September 2025 and today, with China's share of that usage climbing from about 70% to over 80%; OpenCode runs about 95% or more of its inference on Chinese models.
  • Mentions of any open model in new arXiv AI/ML papers rose from 2% in January 2023 to 50% in September 2026, and China's overall share of those mentions now exceeds 40% against the U.S.'s 30%.

Why it matters

This is a rare case of an informal industry newsletter doubling as public testimony: the same numbers the author says he gave Congressional staff are laid out here with sources attached, turning an often abstract "AI race" narrative into measurable facts, downloads, benchmark scores, deployed-traffic share, that the author tracks himself through his own American Truly Open Models project. It also puts a concrete, uncomfortable data point behind the U.S.-China open-model gap: Hugging Face's own security team reached for a Chinese open-weight model during a cyberattack investigation because closed models declined to help.

Who it affects

U.S. policymakers weighing whether to restrict access to Chinese open-weight models; American labs building open weights, including Meta's Llama, Nvidia's Nemotron, the Allen Institute's Olmo and Thinking Machines' Inkling, all shown losing ground; companies already shipping products on Chinese open models, named examples being Harvey, Cursor, DoorDash, Airbnb and Perplexity; and academic researchers, the large majority of whom the author says already run their experiments on Qwen.

How to use it

The practical read for anyone choosing an open-weight model today: Chinese models lead on Hugging Face downloads, on the Artificial Analysis Intelligence Index, and on deployed traffic, over 80% of OpenRouter's open-model usage and about 95% of OpenCode's inference run on Chinese models. The advantage is strongest for agentic coding, per the piece, where Chinese labs are closest to the closed American frontier; for open-ended scientific tasks such as physics or biology, the gap is wider and American alternatives remain more competitive.

How solid is it

The figures come from sources the author says he maintains or can point to directly: Hugging Face's public download counts, the Artificial Analysis Intelligence Index, OpenRouter's published usage statistics, and his own scan of every paper in five arXiv ML categories. That is stronger sourcing than an anonymous estimate, but it remains one analyst's interpretation and estimates for the closed-frontier gaps (2-5 months, 6-9 months, 1-2 months for a hypothetical distillation ban) are explicitly his own approximations, not measured figures. The captured text also cuts off mid-sentence just as the piece turns to "signs of hope" for American models, so the closing argument is not visible here.

Risks and caveats

The source does not name the specific Chinese model Hugging Face used to analyze the cyberattack it references, and it gives no author name anywhere, only first-person references to being invited to brief Congress, so the byline behind this testimony is not established here. Several headline figures are rounded, approximate ranges ("about", "approximately") rather than precise counts, and the claim that fully blocking distillation would add only 1 to 2 months to the American-to-Chinese gap is the author's own modeled estimate, not an observed outcome.