AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Source Summaries

The current balance of power in open models (Nathan Lambert, September 2026)

medium confidence · updated 2026-09-23

Nathan Lambert's expanded prepared remarks for a September 2026 briefing to members of Congress and staff on open-weight models and U.S.-China competition: defines open-weight versus open-source models, reports China's lead in Hugging Face downloads, Artificial Analysis scores and adoption on OpenRouter, OpenCode and arXiv, estimates Chinese open-weight models trail the closed American frontier by 2-5 months and American open-weight models by 6-9, puts distillation's contribution at 1-2 months, and argues for continued U.S. investment in open models and ecosystem-level risk preparation.

Contents

"The current balance of power in open models" is an essay by Nathan Lambert published on his Interconnects newsletter on September 21, 2026, subtitled "The expanded form of a testimony I prepared for Congress." Lambert writes that he was invited to brief a group of members of Congress and staff on open-weight models "in the lens of U.S.-China competition," and publishes his prepared remarks as "a state of the union on open models." It is a position piece: the figures it reports come mostly from tools Lambert maintains himself (the Interconnects dashboard and the ATOM Project), from Artificial Analysis, and from his own scan of arXiv, and its estimates of capability gaps are his own. It follows his July 2026 essays 6 months to live for open models (Nathan Lambert, July 2026) and Kimi K3: The open-weights escalation (Nathan Lambert, July 2026).

Definitions

Lambert defines open language models as models whose weights are publicly available for inspection or downstream use, contrasted with closed models offered only through APIs (he names GPT-4 and Claude Opus 4.5) or through products (ChatGPT, Claude Code). He divides open models into two categories:

  • Open-weight models, the more common form, such as Meta's Llama, Alibaba's Qwen, Google's Gemma and DeepSeek's models. They are governed by licences setting the terms of downstream use and are often accompanied by inference code in libraries such as Transformers, vLLM and SGLang. He writes that Chinese AI companies have been "the clear leader in open-weight models" since about April 2025.
  • Open-source models, which add the complete information needed to reproduce the model: training code and training data. He writes that the most prominent have been built in the United States, led recently by the Allen Institute for AI's Olmo models, which he helped build "in my recent 2.5 years there," with OpenAthena's Marin models and EleutherAI's Pythia models as other examples from American nonprofits.

He holds that every label, including "closed," sits on a spectrum. His example is Nvidia's Nemotron models, which he describes as "far more open than most open-weight models" because Nvidia releases large quantities of training data under permissive licences, but not fully open-source because not all data is released.

Competitive position of American and Chinese open-weight models

Lambert writes that the Chinese models GLM-5.2 and Kimi K3 "have enacted a step change in the commercial viability of open models," crossing a threshold in agentic capability similar to the one he says Anthropic's Claude Code crossed in December 2025.

He dates China's overtaking of American open-weight models, in research and commercial use, to "about 18 months ago." His metric is Hugging Face downloads, where he writes China took the lead in July 2025, "primarily through the success of Alibaba's Qwen models." Since he published the American Truly Open Models (ATOM) Project in August 2025, he reports, China's download lead has grown to about 1.6 billion, "with a total of 3.2B downloads, twice that of America's total," citing the region-level cumulative view of his Interconnects dashboard.

On the Artificial Analysis Intelligence Index, "as of writing this on September 14, 2026," he gives the top three Chinese open models as Z.ai's GLM-5.3 (45), GLM-5.3-Flash (42) and Moonshot AI's Kimi K3 (44), and the leading American ones as Thinking Machines' Inkling and Inkling Small (26 each) and Nvidia's Nemotron 3 Ultra (23). He does not state an index version. A figure caption in the post states that the top American open models are "behind 15 other Chinese made models" on the index. He adds that the top American models were released in June and July 2026 and are updated less often than Chinese ones, that Chinese labs released models scoring above them 2–6 months before the American companies got there (citing GLM-5 and DeepSeek V4 Pro), and that although more American companies, among them Arcee AI, Poolside and IBM, are releasing open models, "they are not rapidly closing this performance gap."

His gap estimates are that Chinese open-weight models are "approximately 2-5 months behind the closed American frontier," and American open-weight models "approximately 6-9 months behind the likes of OpenAI and Anthropic." Chinese labs are closest on tasks with clear user demand such as agentic coding and further behind on open-ended scientific tasks such as physics or biology.

Explanations for Chinese performance

Lambert calls the reasons Chinese labs produce strong models with fewer resources "still an open debate," influenced by work culture (linking his earlier report from inside Chinese labs) and by technical factors. The labs release faster and focus on a slightly narrower distribution of tasks, which he writes flatters them slightly on public benchmarks: because all labs progress steadily, a model finished later tends to score higher. He nonetheless calls the Chinese models "genuinely strong" and "real competition to the American industry," and writes that this competition "will not decrease meaningfully" as closed labs patch the API vulnerabilities that enable distillation.

On distillation, he writes that it is "most impactful in new domains and does not make it trivial to create a universally strong final model," and estimates that if it were fully prevented, for example with know-your-customer tools at Anthropic and OpenAI, the gap from the strongest American models to Chinese open-weight models "would only increase by 1-2 months." In his summary he describes distillation as "an industry standard technique," most prevalent in the Chinese industry, "which has used basic exploits to extract reasoning traces and additional data from American companies' products that are not fully secured," and states that it "does not explain the entire story" of Chinese labs' success.

As a further factor he reports that Chinese labs changed their posture towards buying training data during 2026. Earlier in the year, he writes, top labs including Moonshot AI and Z.ai preferred to build data workflows in-house; by the summer they had begun buying "challenging RL environments for agentic tasks" from established American companies and new Chinese startups.

Risk and regulation

Lambert links the advance of Chinese open-weight models to "growing regulatory uncertainty," citing documented cybersecurity risks from frontier models including the OpenAI–Hugging Face incident (The Hugging Face Incident and the Road Ahead (OpenAI, August 2026)). He calls it a structural challenge that "there are few effective methods for stopping pieces of open software from reaching bad actors," and argues that restricting access to the strongest Chinese open-weight models would set back American businesses. His example is that Hugging Face used a Chinese open-weight model to understand the cyberattack "because closed models would not answer their requests" (Security Incident Disclosure — July 2026 (Hugging Face)). Managing the risks of open-weight models, he writes, "often comes down to ecosystem preparation," linking an essay by Joshua Saxe.

His recommendation is to "continue to enable investment in open models in the US," which he argues is the best path to get ahead of these risks and of relationships in which American companies rely on Chinese-built models, since ownership of open models "allows better coordination and preparation of risks that are global in their nature while accelerating diffusion of AI services throughout the domestic economy."

Adoption

Inference platforms

Lambert reports that OpenRouter, which he describes as "primarily known for trying different open-weight models" and which has shared usage data for top models since January 1, 2025, grew from about 1 trillion tokens processed from open models in a week of September 2025 to about 80 trillion tokens per week at the time of writing, while the Chinese models' share rose from about 70% to over 80% of that usage. The open-source coding agent OpenCode shows, he writes, an inference volume of about 95% or higher on Chinese models. He cautions that these platforms are "the best approximation" available, since much open-model usage runs on platforms that do not disclose per-model breakdowns, naming Together AI and Fireworks AI, and in private enterprise deployments.

Companies building on Chinese models

He lists Harvey, Cursor, DoorDash's use of Kimi models, Airbnb's use of Qwen and Perplexity's use of DeepSeek as companies building AI features on Chinese open-weight models, calls them "the tip of the iceberg" relative to younger Silicon Valley startups seeking low-cost, flexible options, and reports "a growing trend" of American companies entering enterprise agreements with Chinese labs to obtain permission to use their models in products, "a new form of cross-border technology collaboration I have not witnessed in my career."

Academic research

Lambert writes that "to a first order approximation, most of academic research is conducted on Alibaba's Qwen family," and that members of Qwen's leadership he met in China are "very invested in and intentional about this type of adoption." He reports a scan of every paper in five arXiv categories (cs.AI, cs.CL, cs.CV, cs.LG, stat.ML):

PeriodMeasureFigure
January 2023Papers mentioning any open model2%
April–May 2023New AI/ML papers mentioning at least one prominent open model familyabout 2,600 of 12,000
April–May 2023Scanned papers mentioning Llama / a Chinese model~5.5% / ~1%
Fall 2024Papers mentioning Llama / Qwenabout 23% / about 7.5%
September 2026Papers mentioning any open model50%
"Today" (September 2026)Papers mentioning Llama / Qwenabout 21% / 30%
"Today" (body text)Papers mentioning any Chinese / any U.S. open-weight model"over 40%" / 30%
"In recent months" (summary list)AI papers mentioning Chinese / U.S. open-weight models38% / 28%

He concludes that "we clearly have a lot of work to do in order to re-establish the U.S. as the home of AI research."

American models

As "signs of hope" he cites his group's research (arXiv 2604.07190) finding that American models "of comparable capabilities-to-size regions to their Chinese counterparts get adopted at disproportionate rates"; OpenAI's gpt-oss, its first open-weight models since ChatGPT, becoming "one of the most adopted open-weight models of all time"; Google's Gemma 4 models as "some of the only ones ever" to approach the adoption of Qwen's most popular small models; and Nvidia's Nemotron models, which have "modest adoption despite numerous more capable models at the same size point."

Summary claims

The post closes its analysis with six numbered claims:

  1. The gap from open to closed models available to users has been decreasing over three years and "can be estimated as a 2-5 month gap in capabilities."
  2. Open-model usage is "exploding in high-value industries," naming software engineering, legal services and financial services, and inference platforms built mainly on open models (Together, OpenRouter, Fireworks, Baseten) are "the first winners of an open model post-training economy," with fine-tuning APIs such as Thinking Machines' Tinker as another layer. He cites anecdotes of AI-industry staff using GLM-5.3 instead of Claude or GPT for speed, lower prices, customization and privacy.
  3. Chinese companies are "the clear leaders in open weight models"; American labs have been recovering since DeepSeek R1 in 2025, but Chinese labs "regularly are producing notably stronger models."
  4. Distillation "does not explain the entire story," and "the best estimates" put its contribution at 1–2 months of gap reduction.
  5. Chinese models, particularly Qwen, are "a foundational layer" of research, which with "other political factors and the closed nature of leading American AI companies" contributes to "an accelerated decline in America's lead as the preeminent AI research hub."
  6. Open-weight models are reaching capability levels where new risks, such as cybersecurity, "necessitat[e] an ecosystem level response," amid regulatory attention with "massive uncertainty on how policy would be legally enacted," while many researchers rely on open models because closed models "often refuse critical cybersecurity defensive work or biology research."

Conclusions

Lambert writes that Chinese labs "do not appear to be meaningfully challenged" as leaders of the open-weight ecosystem. He describes the capabilities at a few closed labs as "a precursor to what will soon emerge in the open model ecosystem," and open models as "the substrate for everyone else in the world outside of the few true frontier AI labs," a source of "soft power, influence, and potential" for whoever provides them. With "tens of organizations and thousands of people that are within a year of the frontier," he writes that it is "a matter of when, not if" open models cross thresholds that enable new workflows, notes that data on use outside the U.S. and China is lacking, and calls for understanding how to use open intelligence "for good while proactively mitigating the potential harms." He thanks Florian Brand and Kevin Xu for feedback.

Discrepancies noted

The post gives two sets of figures for the share of papers mentioning Chinese and American open-weight models: "over 40%" against 30% in the adoption section, and 38% against 28% "in recent months" in the summary list. The post does not reconcile them, and both are recorded above.

Relation to other assessments

Provenance

Captured from the canonical Interconnects URL on September 21, 2026, with the complete article body; figure images are retained as links, and reader comments and sidebars are omitted. Recorded in New Developments Log/2026-09-21-0830-ai-developments.md, which reported 26 as the score of Inkling only; the post gives 26 for both Inkling and Inkling Small and 23 for Nemotron 3 Ultra.

Relationships

Categories: Source Summaries | us | china | opinion | open-source | competition | frontier-models | safety
What links here (21)