NeoHorse-1-9B is a 9-billion-parameter causal language model from TokenRhythm, post-trained from Qwen3.5-9B as an initial prototype on the path toward recursive self-improvement (RSI), targeting text-based agent harnesses, tool use, coding, and instruction following. Its core innovation is a routing harness that assigns tasks to a heterogeneous model pool, records tool interactions and outcomes, estimates capability demand, and feeds that signal back into shaping the next training mixture — using routing-guided curriculum SFT and routing-guided on-policy distillation to turn execution trajectories into training data, supported by rigorous deduplication, decontamination, and six-dimensional semantic evaluation of training data. This release contains language-model weights only (vision weights excluded, with configuration and tensor keys repackaged for text-only inference) and retains a 262,144-token native context window extensible to 1,010,000. Across a ten-benchmark evaluation against Granite-4.2-8B, Qwen3.5-9B, Ornith-1.5-9B, Gemma-4-12B-it, and the much larger Muse-Glimmer-30B, NeoHorse-1-9B posts the best overall macro average (69.04 vs. 65.60 for its Qwen3.5-9B base, a +3.44 gain), with standout improvements on agentic benchmarks like VitaBench (+11.00), PinchBench (+7.70), and QwenClawBench (+4.69), alongside strong tau2-Bench (90.82) and BFCL v4 (67.43) scores, though instruction-following gains were mixed. It's servable via SGLang or vLLM with Qwen3-style reasoning and tool-call parsers, and is released under the Apache License 2.0.
