NeoHorse-1-9B fine-tunes Qwen3.5-9B with a routing harness for agentic post-training, gaining +3.44 macro average across ten benchmarks.
What is NeoHorse-1-9B? #
NeoHorse-1-9B is a 9-billion-parameter language model released by TokenRhythm, built by post-training Qwen3.5-9B with a technique the team calls routing-guided agentic post-training. It’s designed for tool use, coding, instruction following, and agent harness environments rather than general chat, and the model card reports a 69.04 macro average across ten benchmarks, a 3.44-point gain over the Qwen3.5-9B base model it was derived from.
TL;DR #
- NeoHorse-1-9B is a fine-tune of Qwen3.5-9B released by TokenRhythm under an Apache-2.0 license, aimed squarely at agentic and tool-use workloads rather than open-ended chat.
- The model uses a routing harness that assigns tasks across a pool of models, logs tool interactions and outcomes, and feeds that signal back into the next round of training data selection.
- Reported results show a 69.04 macro average across ten benchmarks versus 65.60 for the Qwen3.5-9B baseline, a gain the model card attributes to routing-guided curriculum SFT and on-policy distillation.
- Benchmark tables place NeoHorse-1-9B ahead of comparably sized models like Granite-4.2-8B and Ornith-1.5-9B on most agentic tasks, and it beats the much larger Muse-Glimmer-30B on several individual benchmarks despite having roughly a third of the parameters.
- The model supports a 262,144-token native context window , extensible up to 1,010,000 tokens, and ships in BF16 safetensors format across four shards.
- TokenRhythm frames this release as an early step toward recursive self-improvement (RSI) , where models evaluated by the harness can be retrained and returned to the same loop, though the current release is described as a prototype rather than a closed self-improving system.
- The data pipeline behind the fine-tune includes deduplication, benchmark decontamination, and semantic labeling of task scenes, goals, and outcomes, which the team says protects against benchmark contamination and low-quality training examples.
Other agents start typing. Remy starts asking. #
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How does the routing harness work? #
The core idea behind NeoHorse-1-9B isn’t a new architecture. It’s a training methodology built around what TokenRhythm calls a routing harness. Instead of fine-tuning on a static dataset and calling it done, the harness sits between a pool of heterogeneous models and a stream of agentic tasks. It assigns each incoming task to a model in the pool, records how the model uses tools and what outcome it produces, and estimates which capabilities are in short supply based on where models struggle.
That capability-gap signal then shapes what goes into the next training mixture. Rather than uniformly sampling training data, the system prioritizes examples that address the specific weaknesses the harness detected. The updated model can be sent back through the harness for another round of evaluation, in principle closing a loop of evaluate, select, and update.
According to the model’s technical report, this loop is currently a prototype: it demonstrates the mechanism across one or a few iterations rather than running indefinitely. The company positions NeoHorse-1-9B as a step toward recursive self-improvement, not a claim that the model is already self-improving in an open-ended way. That distinction matters. Recursive self-improvement is a term with a lot of baggage in AI safety discussions, typically implying a system that keeps upgrading itself without bound. What’s described here is closer to a structured feedback loop between evaluation and data curation, which is a meaningfully narrower and more verifiable claim.
What does agentic post-training actually change in the model? #
Two techniques sit inside the harness: routing-guided curriculum SFT and routing-guided on-policy distillation. Both are ways of turning raw execution trajectories (the sequence of actions, tool calls, and outcomes a model produces while working an agentic task) into supervised training signal.
The “curriculum” part means training examples are sequenced or weighted based on where the routing harness identified capability gaps, rather than treated as a flat, randomly ordered dataset. The “on-policy distillation” part means the model learns from trajectories generated under conditions similar to its own deployment behavior, which tends to produce training signal that’s more directly useful than distillation from a mismatched teacher distribution.
Before any of that training happens, TokenRhythm describes a data-cleaning pipeline: removing exact and near-duplicate examples, decontaminating against evaluation sets so benchmark answers don’t leak into training data, validating structural correctness of agent trajectories, and running six-dimensional semantic evaluation with subscene-level labeling of scene, goal, and outcome. This kind of labeling lets the training pipeline distinguish between a trajectory that technically executed correctly and one that actually achieved the intended goal, which is a distinction generic instruction-tuning datasets often miss.
How does NeoHorse-1-9B compare to other 9B models? #
The model card benchmarks NeoHorse-1-9B against five open-weight peers: Granite-4.2-8B, Qwen3.5-9B (its own base model), Ornith-1.5-9B, Gemma-4-12B-it, and the notably larger Muse-Glimmer-30B. Comparisons run across ten benchmarks grouped by capability, with agentic tasks given particular weight.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
On QwenClawBench, NeoHorse-1-9B scores 48.73, ahead of every listed competitor including the 30B model. On PinchBench it posts 82.25, again the top score in the group, with an 7.70-point gain over its own Qwen3.5-9B base. VitaBench shows an 11-point jump over the base model, though Muse-Glimmer-30B still edges it out in absolute terms on that particular benchmark. On BFCL v4, a widely used function-calling and tool-use benchmark, NeoHorse-1-9B scores 67.43, topping the comparison set including Ornith-1.5-9B’s 65.03.
The pattern across these results is consistent: NeoHorse-1-9B generally outperforms its own base model and other similarly sized 9B-class competitors, and it’s competitive with or better than the much larger 30B model on several tasks despite having a fraction of the parameters. That’s the practical case for routing-guided post-training as a technique, it’s a way to extract more agentic capability from a fixed parameter budget by being deliberate about what the model trains on next, rather than just scaling up.
Is NeoHorse-1-9B worth using for agent workloads? #
For teams building tool-using agents, coding assistants, or task-execution pipelines on a budget that rules out larger models, NeoHorse-1-9B is a reasonable candidate to evaluate. It’s released under Apache-2.0, which permits commercial use without the licensing friction some open model families carry, and its context window (262K tokens native, extensible to over a million) is large enough for long agent trajectories, multi-file codebases, or extended tool-call histories. The caveat is that benchmark gains reported by a model’s own creator should be treated as a starting point, not a final verdict. The comparison set here is chosen by TokenRhythm, and independent third-party evaluation on your specific task distribution is the only way to know if the routing-harness gains generalize to your use case. The model card itself frames this as a prototype on a path toward a bigger goal, which is a reasonable amount of hedging for a 9B release with a narrow, if well-documented, evaluation suite.
Frequently Asked Questions #
What base model is NeoHorse-1-9B built on?
It’s a fine-tune of Qwen/Qwen3.5-9B, post-trained by TokenRhythm using routing-guided agentic post-training. This release contains language-model weights only, repackaged for text-only inference, with no vision components included.
What license does NeoHorse-1-9B use?
Apache-2.0, which allows commercial use, modification, and redistribution with minimal restrictions.
What is recursive self-improvement in this context?
It refers to a loop where a routing harness evaluates models on agentic tasks, identifies capability gaps, and uses that feedback to shape the next round of training data. Updated models can be sent back through the same harness. TokenRhythm describes NeoHorse-1-9B as an early prototype of this loop, not a fully closed or open-ended self-improving system.
How big is NeoHorse-1-9B and what format does it ship in?
It’s approximately 9 billion parameters, released in BF16 precision as safetensors across four model shards, with a total download count over 6,000 on Hugging Face at time of writing.
How does it perform against larger models?
On several agentic benchmarks, including QwenClawBench and PinchBench, NeoHorse-1-9B scores higher than the considerably larger Muse-Glimmer-30B, though the larger model still wins on some individual benchmarks like VitaBench and WorkBuddy Bench.