NeoHorse-1-9B

// Overview

NeoHorse-1-9B is the larger of the two open-weight models TokenRhythm published on 7 September 2026, with the technical report following on 9 September. It is a post-trained model rather than a new pretraining run: TokenRhythm started from Qwen3.5-9B-Base and fine-tuned it on data produced by its own agent infrastructure.

That method is the point of the release. The report — "NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness" (arXiv 2609.08183) — describes routing agent work across a heterogeneous pool of models, then turning the resulting interaction records into training examples for a smaller model. On TokenRhythm's ten-benchmark macro average the 9B model reaches 69.04 against 65.60 for the Qwen3.5-9B base it was trained from.

The gains are concentrated in agent work rather than spread evenly. PinchBench moves from 74.55 to 82.25 and VitaBench from 31.25 to 42.25, while LiveCodeBench v6 (65.14) and IFBench (66.33) are unchanged from the base and IFEval slips slightly, from 89.46 to 89.09. In TokenRhythm's published table the model leads all six columns on QwenClawBench, PinchBench, BFCL v4, tau2-Bench and the overall average, but trails Gemma-4-12B-it and Muse-Glimmer-30B on instruction following.

It keeps the base model's 262,144-token native context, documented as extensible to 1,010,000 tokens, and is served like any Qwen 3.5 derivative: SGLang or vLLM behind an OpenAI-compatible endpoint, with the qwen3 reasoning parser and the qwen3_coder tool-call parser. Weights are Apache-2.0 on Hugging Face and ModelScope, with GGUF and MLX builds for local use.

// Benchmarks

This model's scores

1. tau2-Bench90.82%
2. IFEval89.09%
3. HumanEval98.17%
4. PinchBench82.25%
5. Ten-benchmark average69.04%
6. BFCL v467.43%
7. IFBench66.33%
8. LiveCodeBench v665.14%
9. QwenClawBench48.73%
10. VitaBench42.25%
11. WorkBuddy Bench40.15%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Leads TokenRhythm's published comparison on the ten-benchmark average with 69.04, ahead of Muse-Glimmer-30B (67.86) and Qwen3.5-9B (65.60)
• Strong tool-use and agent-harness scores: 90.82 on tau2-Bench, 82.25 on PinchBench and 67.43 on BFCL v4
• 262,144-token native context, documented as extensible to 1,010,000 tokens
• Apache-2.0 weights on Hugging Face and ModelScope, plus GGUF and MLX quantisations for local runs
• Drops into an existing Qwen 3.5 serving stack — SGLang or vLLM with the qwen3 reasoning and qwen3_coder tool-call parsers

// Best for

• Self-hosted agent harnesses that need tool calling and multi-step task execution on a single node
• Replacing a hosted mid-size model in a tool-heavy pipeline where weights must stay on your own hardware
• Long-context agent runs — browser or repository work — that would be costly to send to a frontier API
• Reproducing or extending the agentic post-training recipe from the NeoHorse-1 report on your own task mix

// How to access

// NeoHorse — every version

The full lineage of the NeoHorse line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.