Intel / article

The current balance of power in open models

Source: interconnects.ai
Author: Nathan Lambert
Date: Published Sep 21, 2026

From the Daily Brief

AI researcher Nathan Lambert published briefing notes prepared for a congressional audience that assess the current balance of power between the United States and China in open weight AI models. The notes draw a clear line between open source, where training code, data, and methodology are also public, and open weight, where only the trained parameters are released, a distinction that shapes what practitioners can legally inspect and build on.

Lambert argues Chinese labs have led open weight releases since roughly April 2025, ahead of most Western labs, which have increasingly kept their strongest models closed while competing on other terms. The briefing lays out the strategic reasoning behind that split. Releasing open weights builds developer mindshare and ecosystem lock in even when it forgoes direct licensing revenue, a tradeoff Chinese labs have leaned into more consistently than their American counterparts.

For engineers deciding what to build on, the practical takeaway is that the strongest open weight options increasingly come from Chinese labs rather than American ones, and that open weight availability itself has become a geopolitical signal, not just a technical or business choice.

Read the 2026-09-22 Brief →

Context

Nathan Lambert's assessment of open models treats research adoption as one measure of influence. In his September 2026 account, he reports scanning papers in five AI and machine-learning categories on arXiv. About 30% mentioned Qwen, compared with roughly 21% mentioning Meta's Llama.

Those figures measure mentions in the scanned literature, not global market share or comparative accuracy. They nevertheless support his argument that Chinese models have become an important foundation for research: influence includes which models other researchers choose to study and build upon.

Key takeaways · AI-distilled

On the Artificial Analysis Intelligence Index (Sept 14, 2026), Z.ai's GLM-5.3 (45) and GLM-5.3-Flash (42) and Moonshot's Kimi K3 (44) beat the top US open models: Thinking Machines' Inkling (26) and Nvidia's Nemotron 3 Ultra (23), per author Nathan Lambert.
OpenRouter's open-model traffic grew from about 1T /week in September 2025 to about 80T tokens/week today, with Chinese models' share of that usage rising from about 70% to over 80% over the same period, Lambert reports.
Scanning arXiv's five top ML categories, Lambert found mentions of any open model rose from 2% of papers in January 2023 to 50% in September 2026, with Chinese models now cited in over 40% of papers versus 30% for the US, reversing Llama's earlier lead.
Lambert estimates that fully blocking of US model outputs, e.g. via know-your-customer checks at OpenAI and Anthropic, would widen the Chinese-to-American open-model gap by only 1-2 months, since distillation isn't the main driver of Chinese labs' progress.
Chinese open-weight models trail the closed US frontier by roughly 2-5 months while open-weight US models trail OpenAI and Anthropic by about 6-9 months; Chinese labs are closest on agentic coding, furthest behind on open-ended science like physics or biology.

Terms in this piece · Glossary

open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.

Read the source www.interconnects.ai