EVENT DOSSIER
On September 8, 2026, KineMotion, in collaboration with Wuwen Xingdong, Tsinghua University, Peking University, and other institutions, officially released its first Agent-Native model, NeoHorse-1. The model comes in two versions: 4B and 9B, and was trained based on Alibaba’s open-source Qwen3.5-4B and Qwen3.5-9B models. During training, KineMotion used the execution trajectories generated by Routing Harness as core data, combined with routing signal estimation tasks to determine training sequences, and updated the model through agent-based supervision and strategy distillation. Evaluation results showed that NeoHorse 4B achieved the highest unweighted average score and surpassed Qwen3.5-9B in five benchmark tests. Its performance improvements were particularly evident in tasks where processes were clear, feedback was observable, and results could be verified; however, the larger model still had advantages in complex state maintenance, long-term debugging, and failure recovery. This release aims to achieve single-round engineering validation through Recursive Self-Improvement (RSI)…
基元律动 NeoHorse-1 联合多家机构发布首个Agent-Native模型
基元律动联合无问芯穹、清华及北大等机构发布首个 Agent-Native 模型 NeoHorse-1，包含 4B 和 9B 两个版本。该模型基于阿里巴巴开源的 Qwen3.5-4B 和 Qwen3.5-9B 进行后训练，利用 Routing Harness 产生的执行轨迹作为核心语料，结合路由信号估计任务能力、安排训练顺序，并通过 Agent 执行监督与策略蒸馏更新模型。评测显示，NeoHorse 4B 在未加权平均分上最高，并在五项基准上超过 Qwen3.5-9B 底座，改善集中在流程清晰、反馈可观察及结果可验证的任务上；更大模型在复杂状态维护、长程调试和失败恢复中仍具优势。此次发布旨在通过递归自我改进（RSI）的单轮工程验证，探索 Harness 驱动的 RSI 路径，为国产开源模型的训练与迭代提供协作基…