TokenRhythm
privateTokenRhythm (基元律动) is a Chinese AI startup founded in April 2026 by Wang Yunhe, the former director of Huawei's Noah's Ark Lab who led its Pangu large-model programme, and Han Kai, formerly a chief researcher at Noah's Ark. Wang announced on March 28, 2026 that he was leaving Huawei after nearly nine years. Shanghai Jiyuan Lüdong Technology was registered on April 8, according to Sina Tech's sources. The website and the company's signed software belong to a Beijing entity, Beijing TokenRhythm Technologies.
Funding:
- Seed round: led by Granite Asia, with CDH VGC, Yuanjing Capital and Infinite Fund.
- August 27, 2026: a round led by HLC Fund (弘晖基金), with Juhe Capital and Shangshi Capital. Coverage gives "tens of millions of US dollars" without making clear whether that is the round or the cumulative total.
- Valuation: Sina Tech reported in June, citing unnamed sources, that a round had closed at a $100M valuation. The company has not confirmed a valuation, so this index records none.
Its business is a routing harness: a layer that picks and combines models for each step of an agent's work, rather than an OpenRouter-style reseller. It uses the traces that routing produces as post-training data, with recursive self-improvement as the stated long-term aim. The pieces:
- OpenSquilla (May 2026): an open-source agent with a built-in model router and about 7,100 GitHub stars. Its July report, "Agentic Routing: The Harness-Native Data Flywheel", lays out the strategy.
- The TokenRhythm API (public beta August 27): a single endpoint over many models. The company claimed 54,000 users and more than 500B tokens a day at launch.
- NeoHorse-1 (September): 4B and 9B open-weight agent models post-trained from Qwen3.5 on routing traces. They are the first turn of that loop. TokenRhythm has not yet pretrained a model of its own.
- InfraRSI (November): a harness that applies the same verify-and-remember loop to the training stack itself. It reports 1.83× faster training on four dense and MoE workloads.
Collaborators on these papers include Yu Wang's group at Tsinghua, Infinigence AI and Peking University. Smaller releases include Claw-SWE-Bench, a benchmark for agent harnesses on SWE-bench-style repair tasks (arXiv 2606.12344). There is also API-TAIL, which detects when an API silently swaps in a different model, using temperature-aligned likelihood tails (arXiv 2610.36801). Finally, there is a survey of agent-system and harness design (arXiv 2606.20683).
Your notes
People
- Yunhe Wang Google Scholar — Founder and CEO (formerly Huawei Noah's Ark Lab (Director; led Pangu); PhD Peking University; BS Xidian University)
- Kai Han Google Scholar — Co-founder and CTO (formerly Huawei Noah's Ark Lab (Chief Researcher); PhD ISCAS; MS Peking University; BS Zhejiang University)
- Yuchuan Tian — Researcher (first author, InfraRSI) (formerly Peking University)
News
- 2026-11-08 InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure (1.83x training speedup, open source) — TokenRhythm
- 2026-09-07 NeoHorse-1: 4B and 9B open-weight agent models post-trained on routing-harness traces — TokenRhythm
- 2026-08-27 基元律动完成数千万美元新一轮融资 TokenRhythm API平台启动公测 (TokenRhythm raises a new round of tens of millions of dollars led by HLC Fund; TokenRhythm API enters public beta) — Shanghai Securities News
- 2026-07-13 Agentic Routing: The Harness-Native Data Flywheel (OpenSquilla technical report) — arXiv
- 2026-06-02 独家|原华为盘古“90后少帅”离职创业,新公司已获1亿美元估值新融资 (Former Huawei Pangu lead Wang Yunhe's new company reportedly raises at a $100M valuation) — Sina Tech