TokenRhythm Raises Millions for Multi-Model Routing; New Research Targets Robot Orchestration and Faster VLA

TokenRhythm's seed round was previously led by Granite Asia, with participation from CDH VGC, Yuanjing Capital and Infinite Fund. CTO Han Kai is a former principal researcher at Huawei Noah's Ark Lab and a longtime collaborator of Wang. Core team members come from Huawei Pangu, major large-model vendors and global cloud companies. The company targets productivity tools, letting developers automatically choose, combine and switch models based on task type, execution stage and cost constraints.

Its first open-source AI agent product, OpenSquilla, has gained more than 6,600 GitHub stars. The company says a domestic multi-model integration solution beat the overseas model Fable 5 on the DRACO complex research task at one-third the cost. TokenRhythm's multi-model API platform, released recently, supports OpenAI- and Claude-compatible protocols and has attracted 54,000 users, with daily token calls exceeding 500 billion. Wang Yunhe said model capability is only part of an agent; organizing models, understanding real-world tasks and letting real-world feedback continuously improve models may be the industry's next challenge.

The routing layer has become a major battleground. In August 2026, Stripe agreed to acquire OpenRouter, the largest large-model routing platform, for about $7.5 billion, after OpenRouter was valued at $1.3 billion in May of the same year. Its backers include a16z, Sequoia Capital, Nvidia and Alphabet, according to the report.

On the research side, a paper released by Huawei Noah's Ark Lab, whose authors are affiliated with several eastern Canadian universities and research institutes, proposes RoboHarness, a system that orchestrates heterogeneous robot policies such as VLA, WAM, RL and TAMP without modifying or retraining them. A Coding Agent makes high-level decisions, assisted by auxiliary skills that quantify whether each policy is suitable for the current sub-task. A Memory Bridge handles state mismatches between policies; in ablations, removing the bridge dropped task success from 86.0% to 60.4%, despite only a modest drop in task progress. This echoes a separate Tsinghua-led system, Harness VLA, which improved success on LIBERO-Pro disturbance benchmarks from 50% to 82.4% by letting a frozen VLA focus on contact-rich manipulation while a harness layer handles resets and retries.

Another IJCAI 2026 paper takes a different approach to the same efficiency problem. The AC²-VLA framework, built on CogACT, uses action state rather than visual complexity to decide how much computation is needed at each step. A unified router controls cognitive caching, visual token pruning and layer skipping. On Google Robot's Visual Matching setting in the SIMPLER benchmark, AC²-VLA reached 76.8% average success across four tasks, surpassing the dense baseline CogACT's 74.8%, while reducing FLOPs to 29.4% of the original and achieving a 1.79x speedup in end-to-end inference. With a cache threshold of 0.2, success rose to 87.1%, which the authors attribute to temporal consistency. First author Yu Wenda said the key training design is self-distillation, allowing the sparse model to mimic the dense teacher's action outputs and backbone features.