Wang Yunhe Unveils First Model Since Starting His Venture
Wang Yunhe, former director of Huawei's Noah's Ark Lab and leader of the Pangu model, has launched his first model, NeoHorse, through his startup Ji Yuan Lü Dong (TokenRhythm). Named NeoHorse-1, the model comes in 4B and 9B parameter versions and is described as "Agent-Native," focusing on abilities crucial for AI agents like tool use, reading environmental feedback, error correction, and task completion.
The model's development was supported by infrastructure from Wu Wen Xin Qiong, with algorithmic research from Tsinghua and Peking University teams. A key innovation is its training data, derived from the execution logs of the company's existing multi-model routing system, OpenSquilla (TokenRhythm's Routing Harness). This system routes tasks between different AI models. The logs provide rich, task-execution trajectories—including successes, failures, and recoveries—which were filtered and used for "Agentic Post-Training."
This post-training, employing techniques like Routing-Guided Curriculum and On-Policy Distillation, significantly improved performance. The 4B model's overall score surpassed that of the 9B base model (Qwen3.5-4B) in comprehensive benchmarks, achieving state-of-the-art results for its size, particularly in structured tasks with clear workflows.
The release positions NeoHorse within TokenRhythm's broader strategy, which includes the OpenSquilla open-source framework, the TokenRhythm API (a model aggregation platform), and enterprise services. The core concept is a potential "flywheel": the routing system generates valuable execution data, which trains better models like NeoHorse; these improved models then enhance the system's efficiency and cost-effectiveness, attracting more usage and generating more training data. This cycle represents an early step toward Recursive Self-Improvement (RSI).
The move addresses a strategic challenge for middleware companies: as major model developers expand into agent infrastructure, TokenRhythm aims to build a moat not just through routing logic, but by converting its unique multi-model operational experience into proprietary model capabilities. The success of this approach in creating a sustainable business remains to be seen.
marsbit09/08 09:27
