Chinese AI startup TokenRhythm raises tens of millions, launches "China's OpenRouter"

Chinese AI infrastructure startup TokenRhythm (基元律动) has announced a new funding round of tens of millions of dollars, led by Honghui Fund, with participation from Juhe Capital and Shangshi Capital, according to a company statement republished by QbitAI. The round follows an earlier seed round led by Granite Asia with CDH VGC, Yuanjing Capital, and Infinity Fund, bringing the company's cumulative fundraising to tens of millions of dollars.

Alongside the funding news, the company launched the public beta of TokenRhythm API, a one-stop multi-model API service benchmarked against OpenRouter for the Chinese market. Developers need a single API key to call multiple models, without separately registering accounts, adapting interfaces, or managing billing.

The platform is compatible with OpenAI and Claude protocols and offers model discovery, intelligent filtering, unified billing, usage statistics, and call logs — effectively a unified, stable entry point between model and inference providers and AI applications.

TokenRhythm said the API has accumulated 54,000 users, processes more than 500 billion tokens per day, and is running proof-of-concept projects with several leading enterprises.

The company does not intend to stop at API aggregation. It is betting on a "Routing Harness" infrastructure layer between models and agent applications. Unlike a model gateway focused on unified access and provider switching, a routing harness enters the agent task execution process, dynamically selecting, switching, coordinating, and aggregating models based on task type, execution stage, cost budget, and real-time state.

Its first product, OpenSquilla, is an open-source AI agent that supports one-click launch across operating systems and has amassed more than 6,600 GitHub stars, 170,000 clones, and over 10,000 real installations. On the PinchBench benchmark, query-level routing cut call costs to about one-ninth of task-level routing at the same accuracy; on the DRACO complex research task benchmark, a multi-model ensemble of domestic models scored above Fable 5 at about one-third the cost.

CEO Wang Yunhe said model capability is only part of agent intelligence: the next stage of competition is not just who owns the strongest single model, but who can better organize models, understand tasks, and let real feedback keep improving models.

What to watch: whether the "China's OpenRouter" positioning can scale beyond the beta as user base and model supply both expand, and whether routing-layer infrastructure becomes a durable moat in a market where multiple models will coexist for a long time.