TokenRhythm · 2026-09-08 · notable
NeoHorse-1 — open 4B and 9B models post-trained by a routing harness
NeoHorse-1 is a pair of Apache-2.0 models, 4B and 9B, that TokenRhythm post-trained by routing agent work across a pool of models and turning the results into training data. The 4B scores 64.87 on the team's ten-benchmark average, up from 58.94.
TokenRhythm routes agent tasks across a pool of models, keeps what worked, and trains 4B and 9B models on it.
Key specs
| 4 b macro average | 64.87 |
|---|---|
| 9 b macro average | 69.04 |
What is it?
NeoHorse-1 is TokenRhythm's attempt to make a small model improve itself from its own agent traces. Two open-weights models ship under Apache-2.0: NeoHorse-1-4B, fine-tuned from Qwen3.5-4B, and NeoHorse-1-9B, from Qwen3.5-9B. Both keep a 262,144-token context window and drop the vision weights for text-only inference. GGUF builds are published alongside the safetensors.
How does it work?
A routing layer sends each task to one of several models of different sizes and records how each one did. Those records become supervised fine-tuning examples, grouped by how hard the task was, and are then distilled back into the smaller model — the paper calls this routing-guided distillation and frames it as a step toward recursive self-improvement. TokenRhythm reports the technique as post-training only; no new pretraining run is claimed.
Why does it matter?
The gains are on agentic work, which is where small models usually fall apart. TokenRhythm reports 64.87 versus 58.94 for the 4B against its Qwen3.5-4B base across ten benchmarks, and 69.04 versus 65.60 for the 9B, including 90.82 on tau2-Bench. Both sizes fit on a single consumer GPU and run under SGLang, vLLM, Transformers, Ollama, llama.cpp and LM Studio.
Who is it for?
developers running local agents
Try it
TokenRhythm/NeoHorse-1-4B on Hugging Face