Lambert's Hill briefing: Chinese open weights out-download US 2-to-1

Share
Lambert's Hill briefing: Chinese open weights out-download US 2-to-1

Nathan Lambert's Hill briefing: Chinese open weights out-download US 2-to-1

Three stories about who sets the terms: the open-weight balance sheet Congress got briefed on privately, the banks trying to write the rules for shopping agents before a regulator does, and the clinics getting frontier medical AI for free.

Chinese open-weight models now hold twice America's cumulative downloads and more than 80% of open-model usage on OpenRouter, according to the remarks Nathan Lambert prepared for a closed-door congressional briefing and published Monday on his Interconnects newsletter. Lambert, until recently the post-training lead at the Allen Institute for AI, puts China at roughly 3.2 billion cumulative Hugging Face downloads against about 1.6 billion for the US — the lead flipped in July 2025 on the back of Qwen. On OpenRouter he measures open-model traffic going from about 1 trillion tokens a week in September 2025 to roughly 80 trillion today, with Chinese models rising from around 70% of that to over 80%.

Read the second number carefully, because it is the one most likely to be repeated wrongly in Washington. That 80% is a share of open-model usage on OpenRouter, not of the platform's total traffic — independent trackers put Chinese models at roughly half of all public token volume there, which is a different and less alarming claim. The downloads figure comes from Lambert's own dashboard; Hugging Face's own spring report corroborates the direction, putting Chinese models at a 41% plurality of downloads with China ahead of the US in monthly volume. The capability gap is the part that should actually worry the US side: Lambert estimates Chinese open weights run two to five months behind the closed American frontier, while the best US open models sit six to nine months back.

His policy ask is conspicuously soft for a briefing titled as testimony. Lambert does not name a bill, an agency, or a number — he argues the best path is to keep enabling investment in US open models, and notes that fully blocking distillation would widen the Chinese gap by only one to two months. We have watched Washington approach this from the other end — Bessent and He put open-weight models on the AI agenda — but the asymmetry Lambert documents is that China's advantage is now distribution, not just capability.


Six banks published a joint principles paper on agentic commerce that says the quiet part in its own text: "Agentic commerce introduces new safety risks, with potential for higher rates of scams, fraud and disputes." The paper, "Building Trust in Agentic Commerce," is co-authored by Bank of America, Capital One, ING, NatWest, Australia's Commonwealth Bank and New Zealand's ASB Bank, and it carries five principles — transparency, safety, privacy and data, choice, and interoperability. The concrete asks are procedural: agents should be identifiable, there should be auditable records of what a consumer instructed and how they authenticated, any party holding potential liability should be able to demand authentication, and issuers and acquirers should get real-time data on the agent's identity and the consumer's intent. The banks' own consumer research is the argument for all of it: "Consumers are unclear if AI agents will act in their interests. They are concerned that AI agents may buy the wrong thing or spend too much – or even worse, lose their money to scams and fraud."

Two things about this document are more important than its contents. The first is that it is voluntary and says so — "Nothing in this paper requires the authors or other industry participants to take any particular action" — which means it is an attempt to set the industry standard before a regulator sets one instead. The second is that it contains no numbers at all: no loss estimate, no fraud rate, no timeline. That is not a reporting gap; there is no figure in the paper. We argued the same point from the other direction when the card networks moved — Visa, Mastercard and Ant want one trust layer for AI agents — and the banks' version of the trust layer is authentication and audit trail, which is cheaper to agree on than liability. A follow-up paper on implementation through protocols and standards is promised.


Anthropic and OpenEvidence are taking the medical AI search tool clinicians already use in the US and Europe and giving it away in about 100 low- and middle-income countries, including Uganda, Angola, Sudan, Haiti and Mongolia. Anthropic supplies the back-end model; OpenEvidence handles regional adaptation, adjusting for local healthcare infrastructure. It is free for providers in the new markets, as it already was for US and European clinicians. The scale argument comes from OpenEvidence chief executive Daniel Nadler: US clinicians consulted the platform 42 million times in August alone, and the company expects several hundred million Americans this year to be treated by a doctor who used it. Anthropic president Daniela Amodei framed the deal as the thing markets would not do on their own: "the market incentives alone would not let it happen without this type of entrepreneurial, philanthropically minded work."

Neither company disclosed financial terms, and Reuters' report does not name which Anthropic model is behind the tool. That omission is the interesting part — Anthropic gets a global health distribution story without publishing an accuracy claim for a clinical setting, in countries where regulatory oversight of medical AI is thinnest. The prior art here is real, though: OpenEvidence was already working with health organizations in Rwanda and Botswana. The gap between free access and verified safety is where this story will be judged, and no independent evaluation was released alongside it.

What to watch: whether the banks' protocol follow-up ever becomes a standard with teeth, and whether a US open-weights policy response lands before the download gap gets wider.

Is the US losing the open-weights race on distribution rather than capability — and would a policy fix even change that? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

AllSpark's Iris beats every open-weight search agent — and says its scaffolding did the work

AllSpark's Iris beats every open-weight search agent — and says its scaffolding did the work

A paper out of AllSpark Research claims the strongest open-weight deep-research agents in their size classes, and the authors' most interesting admission is not a benchmark number: it is that half the story belongs to the harness wrapped around the model. AllSpark released Iris-mini and Iris-pro, two open-weight search agents it says beat every comparable open model on multi-hop browsing benchmarks. The recipe is the unusual part.

DeepSeek shows the machine that trains its agents — and it runs 3 million sandboxes a day

DeepSeek shows the machine that trains its agents — and it runs 3 million sandboxes a day

While Western labs debate whether their agents are safe in theory, DeepSeek just published the factory floor. A new paper, "DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale" (arXiv:2609.22978), lays out the production system behind the lab's agentic reinforcement learning — and founder Liang Wenfeng is the last-listed author of the 162-author submission.

Apple's LensVLM-9B reads text as images at 4.3× compression

Apple's LensVLM-9B reads text as images at 4.3× compression

Apple quietly put a 9B vision-language model on Hugging Face this week that reads documents as pictures instead of tokens — and holds its accuracy down to a quarter of the original text. It is the most interesting long-context idea to come out of a big lab this month, and it landed with almost no announcement. Apple released LensVLM-9B on September 21 — weights under its own research licence, code under its sample-code licence, and a paper that has been sitting on arXiv since May.

Tsinghua open-sources RPent, an agent layer for real robots

Tsinghua open-sources RPent, an agent layer for real robots

Two open-source releases today point the same direction: the interesting work is no longer in the model itself, but in the scaffolding around it. A Tsinghua-led team published the layer that sits between a frontier model and a robot's motors — and a leaderboard to show it lifts a frozen policy. A solo developer, meanwhile, trained a small continual-learning model on a laptop GPU and documented what it forgets. A Tsinghua-led group open-sourced RPent, a framework that turns a general model into…

NVIDIA's 100M-param diarizer tops the speaker-label leaderboard

NVIDIA's 100M-param diarizer tops the speaker-label leaderboard

NVIDIA quietly shipped the missing half of speech AI this week: not what was said, but who said it. Nemotron 3 Diarization is a 100-million-parameter open-weight model that labels up to eight speakers in live or recorded audio, and it arrives ranked first on VoiceArena's initial Diarization-Bench. Across 139 English conversations totalling roughly 22 hours, it scored a 14.72% diarization error rate against 19.3% for the next-best of the 12 systems evaluated — about a 24% relative gap — with…