Former Huawei Pangu head Wang Yunhe starts a business doing RSI, unveiling a 125-billion-parameter agent large model

What happened: Beijing model startup TokenRhythm (基元律动) released NeoHorse-1.5-Flash Preview, an agent-focused large model built on Qwen3.8-Flash through recursive self-improvement (RSI) iteration, and opened it for internal testing. Key numbers: 125 billion total parameters, 6 billion activated. Why it matters: The company uses results from multiple models running the same task as training signals, which it says can also feed its multi-model routing business. Founder and CEO Wang Yunhe formerly led Huawei's Pangu model and Noah's Ark Lab.

Beijing model startup TokenRhythm (基元律动) released NeoHorse-1.5-Flash Preview, an agent-focused large model built on Qwen3.8-Flash through recursive self-improvement (RSI) iteration, and opened it for internal testing.

The company uses results from multiple models running the same task as training signals, which it says can also feed its multi-model routing business. Founder and CEO Wang Yunhe formerly led Huawei's Pangu model and Noah's Ark Lab.