TokenRhythm/NeoHorse-1-9B
Summary
NeoHorse-1-9B is a 9B parameter language model derived from Qwen3.5-9B, fine-tuned for agent tasks and tool use with benchmark performance improvements.
View Cached Full Text
Cached at: 09/18/26, 02:42 AM
TokenRhythm/NeoHorse-1-9B · Hugging Face
Source: https://huggingface.co/TokenRhythm/NeoHorse-1-9B Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness.
NeoHorse-1-9B is a 9B causal language model and an initial prototype on the path towardrecursive self-improvement (RSI). It is post-trained from Qwen3.5-9B for text-based agent harnesses, tool use, coding, and instruction following.
Derived fromQwen/Qwen3.5-9Band fine-tuned by TokenRhythm. This release containslanguage-model weights onlyand is repackaged for text-only inference. Vision weights are not included. Repackaging changes configuration and tensor key names, without changing the fine-tuned tensor values.
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#highlightsHighlights
- **Path toward RSI:**the routing harness 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 can return to the harness, closing a prototype evaluation–selection–update loop; extending this loop across successive iterations is the next step toward RSI.
- **Agentic post-training framework:**the associated research explores routing-guided curriculum SFT and routing-guided on-policy distillation to turn execution trajectories into training signal while preserving execution and harness context around each response.
- **Data quality:**exact and near-duplicate removal, evaluation decontamination, structural validation, six-dimensional semantic evaluation, and subscene-level Scene/Goal/Outcome labeling.
- Broad gains:69.04 macro average across ten benchmarks versus 65.60 for Qwen3.5-9B (+3.44).
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#model-detailsModel Details
PropertyValueModel familyNeoHorse Agent-Native Causal Language ModelParametersApproximately9BBase modelQwen3.5-9BPost-trainingRouting-guided agentic post-trainingInterfaceText input and text outputContext length262,144 natively and extensible up to 1,010,000 tokens.Weight format / precisionSafetensors / BF16
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#evaluationEvaluation
The 9B track compares NeoHorse-1-9B with five representative open-weight baselines: Granite-4.2-8B, Qwen3.5-9B, Ornith-1.5-9B, Gemma-4-12B-it, and Muse-Glimmer-30B. Results cover ten benchmarks and are grouped by capability. Higher is better;Δis NeoHorse-1-9B minus Qwen3.5-9B.Boldandunderlinemark the best and second-best results in each benchmark row, respectively; ties share the same formatting.
BenchmarkGranite-4.2-8BQwen3.5-9BOrnith-1.5-9BGemma-4-12B-itMuse-Glimmer-30BNeoHorse-1-9BΔ vs Qwen3.5-9B🤖 AgenticQwenClawBench
37.01
44.04
47.27
43.53
46.11
48.73
+4.69
WorkBuddy Bench
35.07
39.60
29.29
29.65
45.85
40.15
+0.55
PinchBench
56.93
74.55
68.22
58.89
71.35
82.25
+7.70
VitaBench
23.00
31.25
26.75
36.50
48.50
42.25
+11.00
BFCL v4
52.06
64.88
65.03
62.06
53.74
67.43
+2.55
tau2-Bench
62.28
88.04
83.68
59.37
76.64
90.82
+2.78
💻 CodingHumanEval
96.34
92.68
93.90
100.00
98.17
98.17
+5.49
LiveCodeBench v6
72.00
65.14
47.43
73.14
65.71
65.14
+0.00
📚 Instruction FollowingIFBench
78.00
66.33
40.00
77.67
78.67
66.33
+0.00
IFEval
92.98
89.46
71.35
94.27
93.90
89.09
-0.37
📊 OverallTen-benchmark average
60.57
65.60
57.29
63.51
67.86
69.04
+3.44
**Reported protocol:**SGLang v0.5.17 ·
temperature=1\.0·top\_p=0\.95·top\_k=20·min\_p=0\.0·presence\_penalty=1\.5·repetition\_penalty=1\.0· thinking mode enabled withenable\_thinking=trueandforce\_nonempty\_content=true. QwenClawBench, WorkBuddy Bench, and tau2-Bench use three runs; PinchBench and VitaBench use one run; the remaining benchmarks follow their official protocols. VitaBench uses the DeepSeek-V4-Flash simulator and judge.
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#deploymentDeployment
The examples below are for self-hosted deployment from a downloaded local checkpoint.
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#local-checkpoint-pathLocal checkpoint path
The examples below assume the checkpoint has already been downloaded to local disk. SetMODEL\_PATHto the directory containingconfig\.json, tokenizer files, and model weights.
MODEL_PATH="/path/to/NeoHorse-1-9B"
The OpenAI-compatible requests below use the server’s\-\-served\-model\-name(for example,neohorse\-1\-9b), not the filesystem path.
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#sglangSGLang
The technical report uses SGLang v0.5.17.
pip install "sglang==0.5.17"
MODEL_PATH="/path/to/NeoHorse-1-9B"
python3 -m sglang.launch_server \
--model-path "$MODEL_PATH" \
--served-model-name neohorse-1-9b \
--host 0.0.0.0 \
--port 30000 \
--context-length 262144 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder
Send an OpenAI-compatible request after the server starts:
curl http://localhost:30000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"neohorse-1-9b","messages":[{"role":"user","content":"Write a Python function that returns the first n Fibonacci numbers."}],"max_tokens":512}'
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#vllmvLLM
pip install -U vllm
MODEL_PATH="/path/to/NeoHorse-1-9B"
vllm serve "$MODEL_PATH" \
--served-model-name neohorse-1-9b \
--host 0.0.0.0 \
--port 8000 \
--max-model-len 262144 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder
The server exposes an OpenAI-compatible/v1/chat/completionsendpoint. Send a request after the server starts:
curl http://localhost:8000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"neohorse-1-9b","messages":[{"role":"user","content":"Write a Python function that returns the first n Fibonacci numbers."}],"max_tokens":512}'
The example uses the configured 262,144-token context limit. Actual capacity depends on GPU memory and serving settings; reduce the context limit if needed.
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#licenseLicense
NeoHorse-1-9B is released under theApache License 2.0.
The upstream model isQwen/Qwen3.5-9B. Its original copyright notice, Copyright 2026 Alibaba Cloud, is retained in the license file. TokenRhythm has modified the model through fine-tuning and repackaging for text-only inference. Modification notices are included in this model card and the released configuration, weight index, and Safetensors metadata.
https://huggingface.co/TokenRhythm/NeoHorse-1-9B#citationCitation
@misc{neohorse2026,
title = {NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness},
author = {NeoHorse Team},
year = {2026},
howpublished = {arXiv preprint},
eprint = {2609.08183},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2609.08183}
}
For questions or issue reports, use theNeoHorse project repository.
Similar Articles
TokenRhythm/NeoHorse-1-4B
NeoHorse-1-4B is a 4B parameter language model fine-tuned from Qwen3.5-4B for agentic tasks, using routing-guided post-training to enable recursive self-improvement.
@KyleHessling1: Hello again, everyone! We've got another really fun 9b, this one specifically trained for tool calling and agentic codi…
A new 9B fine-tuned model called Qwopus3.5-9B-Coder is released, optimized for tool calling and agentic coding workflows, achieving strong SWE-bench and HermesAgent-20 scores while running on affordable hardware.
sensenova/SenseNova-Loop-9B
SenseTime's SenseNova team officially releases SenseNova-Loop-9B, an open-source ~9B parameter model that reuses shared middle layers via a looped architecture with manifold-constrained hyper-connections (mHC) to gain deeper computation with fewer parameters, targeting complex reasoning, coding, tool use, and multimodal tasks, with AWQ and GGUF quantized weights now available.
empero-ai/Qwythos-27B-v1
Empero releases Qwythos-27B-v1, an open-weight reasoning model based on Qwen3.5-27B that preserves native multi-token prediction, the full vision tower, and a 1,048,576-token context window. It demonstrates strong agentic terminal performance and improved closed-book reasoning over its 9B sibling.
Qwen/Qwen3.6-27B-FP8
Alibaba releases Qwen3.6-27B-FP8, a 27B FP8-quantized model with strong agentic coding and reasoning benchmarks, now available on Hugging Face.
