BeijingTokenRhythm says its new model gained 22% on agent tests.
Founder Wang Yunhe previously led Huawei’s Noah’s Ark Lab and its Pangu model.
The model, built on Alibaba’s Qwen3.8-Flash, trains on gaps between its answers and fused results from several frontier models.
Its computing and maths score fell 11%, from 59.51 to 52.96, in the company’s own results.
The preview is in beta testing, and every score so far comes from TokenRhythm itself.
How each outlet framed it
- Zhidx leans favorable
- reports the model's gains on Agents' Last Exam (+22.4%) and its drop in computation and math (−11%), a regression the release does not dwell on
Sources: Zhidx