Ex-Huawei Noah's Ark Lead's Startup Ships NeoHorse: Baking Multi-Model Harness Experience Into a 4B Agent Model

量子位 · wechat · 2026-09-08

Jiyuan Lvdong, founded by former Huawei Noah's Ark Lab director and Pangu model lead Wang Yunhe, released its first Agent-Native model NeoHorse, with compute and infra from Infinigence and algorithm research from Tsinghua and PKU teams. NeoHorse-1 comes in 4B and 9B, targeting tool calling, environment feedback reading, error recovery, and path adjustment. After Agentic Post-Training, the 4B (macro average 58.94→64.87) beats its base Qwen3.5-4B across 10 benchmarks, reaching same-scale SOTA.

Core idea: train on Harness execution logs

Business stack and the RSI narrative

The company has four layers: open-source OpenSquilla, TokenRhythm API (a Chinese OpenRouter analog), enterprise services, and now NeoHorse—validating a flywheel from execution experience to training data to model capability. The company frames it as a single-round engineering validation toward RSI (Data-RSI + Model-RSI), though reward and signal design remain human-set and multi-generation gains are unproven. Key risk: DeepSeek, Qwen and others are vertically integrating models with Harness/Agent infrastructure, squeezing the middleware layer.

Original post →

More from Companies & People

Companies & People channel →