TokenRhythm and Partners Release NeoHorse-1 Agent-Native Model Series

A project-scheduling test cited in the technical report illustrates the problem the project aimed to address: a 4B base model located files in the working directory but missed an email containing the latest dependency constraints. It then generated a plan based on outdated information and wrote files to the wrong location.

TokenRhythm was founded by Wang Yunhe, former director of Huawei's Noah's Ark Laboratory and former lead of the Pangu large model. The company earlier developed OpenSquilla, an open-source routing harness system that selects and organizes different models during agent execution. NeoHorse extends that technical route into model training.

According to the technical report, the training corpus is centered on execution trajectories produced by routing harnesses including OpenSquilla, supplemented by public data. These trajectories preserve capability-demand prediction, routing choices, model responses, tool calls and environmental feedback. After integrity checks and quality evaluations such as goal completion, evidence consistency and error recovery, the trajectories are used for post-training.

The team used routing signals to estimate the abilities required by tasks and to arrange training order, then updated the model through agent execution supervision and on-policy distillation. In on-policy distillation, a teacher model gives guidance targeted at what the student model actually generates.

The report evaluated the models on 11 benchmarks covering agent execution, tool interaction, coding and instruction following. Among the 4B-level models listed in the report, NeoHorse-1 4B achieved the highest unweighted average score, and it surpassed the Qwen3.5-9B base model on five benchmarks. The gains were concentrated in tasks with clear procedures, observable feedback and verifiable results, while larger models retained an edge in complex state maintenance, long-horizon debugging and failure recovery.

NeoHorse's release is described as a single-round engineering validation of recursive self-improvement: an agent executes inside a routing harness, evaluation exposes capability gaps, training choices are adjusted, the model is updated and then returned to the harness.

Wuwen Xinqiong, an AI-native infrastructure service provider in China, supplied infrastructure support and infrastructure-optimization technologies, combining algorithm and chip-hardware co-optimization to improve training efficiency and reduce cost. Tsinghua University and Peking University teams participated in algorithm and training-method innovations, aiming to increase the capability return on training investment and push toward a steeper training scaling curve.

NeoHorse is post-trained on Alibaba's open-source Qwen3.5-4B and Qwen3.5-9B. The collaboration links the open-source model foundation, algorithm research, computing infrastructure and agent execution feedback, and it explores training and iteration paths for domestic open-source models while promoting industry-academic cooperation.

For TokenRhythm, NeoHorse provides model-level evidence for its technical vision at this stage. The collaboration also creates a foundation for further iteration, testing whether open-source models can absorb task-execution experience and whether harnesses and models can improve one another in a cycle.