Ex-Pangu chief's TokenRhythm raises again as model routing turns into a land grab
The money is chasing the layer between models and agents. And in robotics, two new papers argue the same lesson from the other direction: stop waiting for one model to rule them all.
TokenRhythm, the AI infrastructure startup founded by former Huawei Noah's Ark lab director Wang Yunhe, has closed another funding round, bringing its total raise into the tens of millions of dollars, according to an exclusive from Leiphone. Hongzhou Fund led the new round, with Juhe Capital and Shangshi Capital participating; the seed was led by Granite Asia with CDH VGC, Yuanjing Capital and Wuqi Fund also in. The company did not disclose exact figures beyond "tens of millions of dollars" and a valuation described only as being in the hundreds of millions.
Why it matters: the bet is that the interesting infrastructure no longer sits inside any single model but in the routing harness above them — picking, switching, and sequencing models per task stage so agents stay effective while costs stay contained. The timing is pointed: Stripe reportedly paid about $7.5 billion for OpenRouter this month, a platform valued at $1.3 billion barely three months earlier, which tells you how fast capital has decided the router is the new traffic entry point for AI. TokenRhythm claims its aggregation API already counts 54,000 users and more than 500 billion tokens routed per day, and says its open-source agent product OpenSquilla (6,600-plus GitHub stars) beat Fable 5 on DRACO complex-research tasks at roughly a third of the cost using a multi-model setup. Those are company-reported numbers — treat them as claims, not audited facts — but the direction matches what the market just paid for. Next step per the company: an agent-native model trained on real routing traffic. We covered Stripe's OpenRouter buy earlier this month — Stripe closes $7B+ deal for AI gateway OpenRouter.
Huawei Noah's Ark researchers have a paper arguing robots don't need a universal model yet — they need a dispatcher. RoboHarness wraps independently built control systems (VLA models, reinforcement-learning policies, task-and-motion planners) as callable skills, then lets a coding agent decide which one handles each piece of a long-horizon task. The hard part isn't choosing policies, it's handing control between them: each policy was trained on different data, so where one stops may be territory the next has never seen. A Memory Bridge module retrieves similar past trajectories, estimates the incoming policy's comfort zone, and steers the robot there before the handoff. Removing it drops full-task success from 86% to 60.4%, the authors report — most failures were tasks dying at the baton pass, not at any single policy. Tested across three public benchmarks, 500 custom tasks and 135 real-robot runs, with no retraining of any underlying policy.
Separately, an IJCAI 2026 paper attacks the other reason robot models feel unusable: they're too slow. AC²-VLA makes the robot's own action state decide how much computation each moment deserves — caching vision-language outputs when the arm's state is stable, pruning irrelevant visual tokens, and skipping layers when deep reasoning isn't needed, all governed by one learned router. Existing efficiency work leans on visual complexity, but the authors point out the mismatch: when the arm nears a delicate grasp, the camera frame looks simple yet demands full compute. Trained via self-distillation against the dense model, the framework cuts FLOPs to 29.4% while running 1.79x faster with success rates intact on the SIMPLER benchmark — and at one cache setting it actually beat the dense baseline by 12.3 points, apparently because reusing features smooths jittery per-frame decisions. It builds on CogACT, and the authors say the distillation recipe transfers to other backbones.
Both papers share one thesis worth watching: orchestration data — records of handoffs, fallbacks and recoveries — may itself be the training data for whatever universal robot model eventually arrives.
If you were building an agent stack today, would you route between many models or commit to one? Tell us in the comments.
Sources: Leiphone · Sina Finance · RoboHarness paper (arXiv) · AC²-VLA paper (arXiv)
