TokenRhythm's NeoHorse-1 Beats Its Base Model by 6 Points on Agentic Benchmarks

A small Chinese lab called TokenRhythm has quietly dropped a 4B parameter model that beats its base by nearly 6 points on average across ten agentic, coding, and instruction-following benchmarks. The release frames itself as a prototype toward recursive self-improvement, where the model's own agentic behavior in a routing harness feeds back into its next training run.

NeoHorse-1-4B is a text-only fine-tune of Qwen3.5-4B released under Apache 2.0. It ships in BF16 safetensors with a GitHub repo, a technical report, and ready-to-run recipes for both SGLang and vLLM.

How the routing harness works

The central design is what the team calls a routing harness. It assigns tasks to a heterogeneous model pool, records tool interactions and outcomes, estimates capability demand, and uses capability-level feedback to shape the next training mixture. Updated models then re-enter the harness, closing an evaluation, selection, and update loop that the authors treat as a first step toward recursive self-improvement.

Concretely, the post-training pipeline combines two techniques: routing-guided curriculum SFT and routing-guided on-policy distillation. Together they convert execution trajectories into training signal while preserving execution and harness context around each response. Think of it as harvesting the model's own agent runs, filtering them by where the router says the fleet is weakest, and feeding those filtered traces back as supervised examples with the surrounding tool state intact.