NeoHorse-1 is a family of agent-native models designed to investigate recursive self-improvement through agentic post-training. It uses a heterogeneous model pool with intelligent routing, recording each user turn’s predicted capability demand, selected service tier, and interaction. Read original