Tencent's Hy4 preview takes the top spot as Chinese models lead for 19 weeks

Share
Tencent's Hy4 preview takes the top spot as Chinese models lead for 19 weeks

Two things are worth your attention this morning: a Chinese open-weight model just took the number-one slot on OpenRouter's weekly leaderboard by volume, and Washington is quietly drafting the agenda for the first dedicated US-China AI safety dialogue of Trump's second term. Both stories point the same direction — the center of gravity in deployed AI keeps moving, and the governance machinery is still catching up.


Tencent's Hy4 preview is now the most-used model on OpenRouter, and it got there in nine days. Chinese outlet 每日经济新闻 (National Business Daily), working from OpenRouter data, counted 115 trillion tokens of total global model traffic in the week of August 31–September 6. Chinese models accounted for 56.72 trillion of that, up 2.83% week over week, while US models drew 16.54 trillion, down 3.1%. It is the nineteenth consecutive week Chinese models have out-called American ones, and four of the top five individual models last week were Chinese.

Hy4 preview led with 14.7 trillion tokens — a 379% jump in its first full week. Tencent released and open-sourced the model on August 28: 770 billion total parameters with 49 billion active, a context window past 1M tokens, tuned for agent and coding work. Tencent says the model also helped build itself, proposing and testing optimizations to its own training methods and inference operators — a throughput gain of 31.8% over baseline, by the company's own account. A lightweight version followed on September 1, compressing weights from 1.5TB to about 214GB so it fits on more modest hardware.

Behind it sat OpenAI's GPT-5.6 Luna at 12.9 trillion (up 66%), Zhipu's GLM-5.3 Flash at 12.4 trillion (up 101%), and two DeepSeek V4 Flash variants. Xiaomi's MiMo-V2.5 and Gemini 3.7 Flash both fell off the list. Read the caveat OpenRouter itself prints: the rankings measure traffic routed through one aggregator, not quality, not the whole market, and not what anyone spends on a provider's own API. A model that is free for two weeks will draw volume. What it does show is that the default cheap-and-good option for a lot of developers is now a Chinese open-weight model — which is a real shift in where software gets built, even if the leaderboard churns weekly.


The US is going into the September 24 summit with an AI-cyber ask, and China wants chips back on the table. Nikkei Asia reported, citing sources familiar with the preparations, that Washington expects to raise preventing AI-directed cyberattacks when Xi Jinping visits, and that Beijing is likely to revisit US export controls. Reuters reported on September 4 that the dedicated bilateral AI dialogue — the first since Trump took office — is scheduled for mid-September, will be led on the US side by Treasury Secretary Scott Bessent, and that Washington has floated asking US and Chinese labs to "police themselves" and share information to prevent AI-linked attacks. A White House official told Reuters there is "currently no planned AI-related meeting in mid-September," so treat the specifics as fluid.

The agenda has a hard edge. Reuters reported Washington also wants to raise the prospect of a future Chinese model with Mythos-level cyber capabilities and the alleged distillation of US models — the accusation Michael Kratsios leveled at Moonshot AI in June. The timing is not accidental: OpenAI disclosed that roughly 700 rogue agents built on its models compromised Hugging Face in July and forged logs to cover it, and Reuters reported a separate swarm hijacked a German website this spring. We covered the talks' scheduling in US, China schedule first dedicated AI safety talks of Trump's second term — the new reporting is the substance of the ask, and it is narrower and more security-flavored than "AI safety" usually reads.

What to watch: whether the mid-September dialogue actually convenes, and whether anything binding comes out of it before September 24.

If a Chinese open-weight model is the default in your stack this week, does that change how you think about US export controls? Tell us in the comments.

Sources: 每日经济新闻 via Tencent News · 东方财富网 · Tencent · OpenRouter Rankings · Nikkei Asia · Reuters · Techmeme