Wang Yunhe unveils NeoHorse, first startup model with agentic post-training insights

NeoHorse's development received computing and infrastructure support from Infinigence and algorithm and training support from Tsinghua University and Peking University teams. Its core training data combines execution signals generated by JiYuanLvDong's Routing Harness with public data, according to a technical report. During a task, the harness logs which capability was assigned, which model was selected, what environment output occurred, and where a failure or recovery happened. That data supply includes both successful and failed trajectories, allowing the model to learn not only answers but also real paths across models.

To turn logs into useful training data, each trajectory went through structural checks and was evaluated on six dimensions, such as whether the user goal was met, whether instructions were followed, whether tools were used reasonably, whether conclusions had evidence, and whether error recovery was appropriate. The company then applied a Routing-Guided Curriculum, using routing signals to order samples from low- to high-complexity tasks. In addition, an On-Policy Distillation step made teacher models guide student models on the actual steps taken rather than a pre-set set of errors.

In evaluations covering harness agent, tool use, code and instruction following, the NeoHorse-1-4B macro-average score rose from 58.94 to 64.87 after agentic post-training, surpassing its base model Qwen3.5-4B and reaching state-of-the-art for models of its size. The improvement was concentrated in tasks with clear workflows and verifiable outcomes. For example, a base model failed to read an email about updated dependencies when scheduling a project, but the post-trained model read the new evidence, recalculated the schedule and saved results to the right location.

JiYuanLvDong said the model does not mean it is abandoning its Routing Harness to become just a traditional model company. Instead, the startup sees a flywheel connecting its open-source OpenSquilla, its unified API TokenRhythm, and its custom models. Experience from routing and task execution can be filtered and used for next-round training, while updated models are placed back into the harness to improve cost and response performance. The company views this as a limited version of recursive self-improvement, with Data-RSI collecting structured execution records and Model-RSI adjusting training distribution based on evaluation results. A key motivation is cost efficiency: if NeoHorse's 4B version can reliably handle a set of frequent agent tasks, it reduces the need to call larger, more expensive models.