Open-sourced in late August, flipped on the router layer by early September. Per National Business Daily’s read of OpenRouter’s latest weekly stats, global model traffic ran about 115 trillion tokens for the week of 31 August–6 September. Tencent’s Hunyuan Hy4 preview took first place with 14.7 trillion weekly tokens—up 379% week over week.
The structure is sharper than the crown. Among listed models, Chinese systems did 56.72 trillion tokens for the week versus 16.54 trillion for U.S. models (down 3.1%). China has now led that slice for nineteen straight weeks. Four of the global top five are Chinese—Hy4, Zhipu GLM-5.3 Flash, and two DeepSeek-V4-Flash builds—with GPT-5.6 Luna in second.
[1][2]Hy4 isn’t selling a parameter poster
Tencent’s 28 August release is specific: 770B total / 49B active, context past 1M, aimed at coding, office, and research work that has to finish. Weights are open; access also runs through WorkBuddy, CodeBuddy, Yuanbao, ima, plus Tencent Cloud TokenHub and OpenRouter. On 1 September a lite cut compressed roughly 1.5TB of weights to about 214GB while keeping long-doc and multi-turn retrieval nearly level with the full model—carpeting for local and edge installs, not another flagship splash page.
OpenRouter’s sticker is frank too: about $0.834 / $2.501 per million in/out. The traffic spike therefore looks more like “defaults on the router got rewritten” than “another open-source fest.”
[2][1]Who the board is punishing
The same week, Xiaomi’s MiMo-V2.5 and Gemini 3.7 Flash fell off the board; MiniMax M3 climbed back to sixth on a timed free tier. Token leaderboards favor models that are cheap, stable, and easy to wire into agent pipelines. They don’t answer “who is closest to AGI.” They answer “where did developers park this week’s token budget.”
For Chinese labs, that is a different soft power: not Arena gold, but becoming the default upstream on overseas keychains. Closed U.S. flagships can keep winning evals while open weights and mid-price APIs chew the invoice share.
[1]Hy4 preview is optimized for agent, coding, and productivity scenarios—finishing complex tasks in real business workflows.
Take: don’t treat OpenRouter as a census of Earth
OpenRouter is a developer-heavy, routable-API slice. It misses Yuanbao chats inside WeChat and most private deployments. Declaring “China AI leads everywhere” from it overreaches; pretending it doesn’t exist is worse—because it measures preference for cross-border purchasable compute, which is the hardest inch of commercialization.
Hy4’s 379% looks like a cold-start win: open weights, product surfaces, and third-party routing opened together. The next test isn’t whether it keeps the crown in week two; it’s whether the 214GB lite cut turns into real local-agent installs. Traffic can spike overnight. Stickiness is repaid slowly in toolchains and cache-hit rates.
[1][2]