The current balance of power in open models

What it claims

Chinese open-weight models lead American ones on downloads, benchmark scores, inference usage and academic mentions, and that lead is not narrowing. The page puts timelines on it. Chinese open-weight models trail the closed American frontier by about 2-5 months. American open-weight models trail by about 6-9 months. The gap is smallest in agentic coding and largest in open-ended science such as physics and biology.

On distillation, the author estimates that fully blocking it, for example with KYC at OpenAI and Anthropic, would widen the gap by only 1-2 months. The stated reasons the Chinese labs do well are:

• they release more often, so each model is a later snapshot of steadily improving training;
• they target a narrower task distribution, which slightly flatters public benchmarks;
• since mid-2026 they have been buying agentic RL environments from US and Chinese vendors instead of building all data in-house.

Evidence

Benchmark: Artificial Analysis Intelligence Index, as of 14 Sep 2026

• The American models were released in June-July 2026.
• Chinese models such as GLM-5 and DeepSeek V4 Pro passed those American scores 2-6 months earlier.
• New entrants Arcee, Poolside and IBM are "not rapidly closing" the gap.

Hugging Face downloads

• China took the lead in July 2025, mainly through Qwen.
• China's lead is now about 1.6B downloads, with 3.2B total, twice America's total.

OpenRouter inference

• Open-model volume grew from about 1T tokens/week (September 2025) to about 80T tokens/week now.
• The Chinese share of that volume rose from about 70% to over 80%.
• The OpenCode coding agent shows about 95% or more of its inference on Chinese models.

arXiv mentions (scan of cs.AI, cs.CL, cs.CV, cs.LG and stat.ML)

• Papers mentioning any open model rose from 2% in January 2023 to 50% in September 2026.
• April-May 2023: about 2,600 of 12,000 papers mentioned a prominent open family. Llama was at about 5.5%, Chinese models at about 1%.
• Fall 2024: Llama about 23%, Qwen about 7.5%.
• Now: Qwen 30%, Llama 21%. Any Chinese model is mentioned in over 40% of papers, any US model in about 30%.

Named adopters of Chinese models

• Harvey and Cursor (models not specified on the page)
• DoorDash (Kimi), Airbnb (Qwen), Perplexity (DeepSeek)
• The author also reports US companies signing enterprise licensing agreements directly with Chinese labs.

Counter-signal for US models: US models that match Chinese models' capability for their size get adopted "at disproportionate rates." Examples given:

• gpt-oss is among the most adopted open-weight models ever.
• Gemma 4 matches the adoption of Qwen's popular small models.
• Nemotron gets modest adoption despite more capable models existing at its size.

Openness spectrum: Nemotron releases large parts of its training data under permissive licences but is not fully open-source. Olmo, Marin and Pythia are cited as the fully reproducible open-source models, all American.

The underlying data is in the Interconnects Dashboard and the ATOM Project trackers. The page links to them but does not reproduce them.

Limits

• Self-reported figures. These are the author's prepared remarks for a policy audience. The figures come from the author's own trackers and arXiv scan, and the scan method (keyword matching, how model families are defined) is not described.
• Inconsistent numbers on the page.
• Academic share is given as over 40% vs 30% in one place and 38% vs 28% "in recent months" in another.
• The page names both GLM-5.2 and GLM-5.3 as current leaders.
• Uncertain estimates. The 2-5 month, 6-9 month and 1-2 month distillation figures are the author's estimates with no stated method.
• Incomplete usage data. OpenRouter and OpenCode are proxies only. Together AI, Fireworks and private enterprise deployments do not publish per-model breakdowns. The author says there is little data on usage outside the US and China.
• Benchmark bias. The author concedes that Chinese labs' narrower task focus slightly inflates their public benchmark results.
• Safety and policy. Restricting Chinese weights would mainly hurt US businesses, according to the author. They cite Hugging Face using a Chinese open model to analyse a cyberattack after closed models refused the requests. No concrete policy mechanism is proposed.