Modelpublic

TokenRhythm/NeoHorse-1-9B-MLX

sourceHugging Faceapache-2.0updated 14d agoView on Hugging Face

2likes582downloads

Model Card

MLX local inference

This is the unquantized BF16 MLX version of NeoHorse-1-9B for Apple Silicon. Converted from the original BF16 weights with MLX-LM. No weight quantization is applied; MLX-LM adapts tensor names/layouts and normalization representation for its runtime. Benchmark scores below refer to the original model, not a separate evaluation of this MLX version.

bash

pip install "mlx-lm>=0.31.3"
mlx_lm.chat --model TokenRhythm/NeoHorse-1-9B-MLX

The model downloads automatically from Hugging Face. The original chat template is preserved. See Deployment for local checkpoints, the chat API, and tool calling.

<div align="center"> <h1>NeoHorse-1-9B</h1> <p><b>Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness.</b></p> </div>

<p align="center"> <a href="https://arxiv.org/abs/2609.08183"><b>Technical Report</b></a> </p>

NeoHorse-1-9B is a 9B causal language model and an initial prototype on the path toward recursive self-improvement (RSI). It is post-trained from Qwen3.5-9B for text-based agent harnesses, tool use, coding, and instruction following.

Derived from Qwen/Qwen3.5-9B and fine-tuned by TokenRhythm. The source checkpoint was repackaged for text-only inference. This repository contains language-model weights only, converted to MLX BF16 without weight quantization.

Highlights

—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).

Model Details

<div style="width:100%;max-width:none;margin:16px 0;padding:0;overflow-x:auto"> <table class="vl-table model-table" width="100%" style="display:table;width:100%;min-width:100%;table-layout:fixed;border-collapse:collapse;font-size:13px"> <thead><tr> <th style="padding:9px 10px;text-align:left;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Property</th> <th style="padding:9px 10px;text-align:left;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Value</th> </tr></thead><tbody> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Model family</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)">NeoHorse Agent-Native Causal Language Model</td> </tr> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Parameters</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)">Approximately <strong>9B</strong></td> </tr> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Base model</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)"><a href="https://huggingface.co/Qwen/Qwen3.5-9B">Qwen3.5-9B</a></td> </tr> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Post-training</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)">Routing-guided agentic post-training</td> </tr> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Interface</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)">Text input and text output</td> </tr> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Context length</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)">262,144 natively and extensible up to 1,010,000 tokens.</td> </tr> <tr> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16);font-weight:600">Weight format / precision</td> <td style="padding:9px 10px;border-bottom:1px solid rgba(249,115,22,.16)">MLX Safetensors / BF16 (unquantized)</td> </tr> </tbody></table> </div>

Evaluation

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; Δ is NeoHorse-1-9B minus Qwen3.5-9B. Bold and <u>underline</u> mark the best and second-best results in each benchmark row, respectively; ties share the same formatting.

<div style="overflow-x:auto"> <table class="vl-table" width="100%" style="display:table;width:100%;min-width:100%;border-collapse:collapse;table-layout:fixed;font-size:13px"> <thead><tr><th style="padding:9px 8px;text-align:left;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Benchmark</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Granite-4.2-8B</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Qwen3.5-9B</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Ornith-1.5-9B</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Gemma-4-12B-it</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.10)">Muse-Glimmer-30B</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.18)">NeoHorse-1-9B</th> <th style="padding:9px 8px;text-align:center;border-bottom:2px solid #f97316;color:#c2410c;background:rgba(249,115,22,.18)">Δ vs Qwen3.5-9B</th></tr></thead><tbody> <tr><td class="benchmark-capability" colspan="8" style="padding:10px 8px;font-weight:700;color:#c2410c;background:rgba(249,115,22,.10);border-top:2px solid #f97316">🤖 Agentic</td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">QwenClawBench</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">37.01</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">44.04</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>47.27</u></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">43.53</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">46.11</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><strong>48.73</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+4.69</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">WorkBuddy Bench</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">35.07</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">39.60</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">29.29</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">29.65</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><strong>45.85</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><u>40.15</u></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+0.55</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">PinchBench</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">56.93</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>74.55</u></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">68.22</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">58.89</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">71.35</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><strong>82.25</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+7.70</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">VitaBench</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">23.00</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">31.25</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">26.75</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">36.50</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><strong>48.50</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><u>42.25</u></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+11.00</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">BFCL v4</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">52.06</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">64.88</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>65.03</u></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">62.06</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">53.74</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><strong>67.43</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+2.55</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">tau2-Bench</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">62.28</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>88.04</u></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">83.68</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">59.37</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">76.64</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><strong>90.82</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+2.78</span></div></td></tr> <tr><td class="benchmark-capability" colspan="8" style="padding:10px 8px;font-weight:700;color:#c2410c;background:rgba(249,115,22,.10);border-top:2px solid #f97316">💻 Coding</td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">HumanEval</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">96.34</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">92.68</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">93.90</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><strong>100.00</strong></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>98.17</u></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><u>98.17</u></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+5.49</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">LiveCodeBench v6</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>72.00</u></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">65.14</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">47.43</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><strong>73.14</strong></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">65.71</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value">65.14</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+0.00</span></div></td></tr> <tr><td class="benchmark-capability" colspan="8" style="padding:10px 8px;font-weight:700;color:#c2410c;background:rgba(249,115,22,.10);border-top:2px solid #f97316">📚 Instruction Following</td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">IFBench</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>78.00</u></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">66.33</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">40.00</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">77.67</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><strong>78.67</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value">66.33</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+0.00</span></div></td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">IFEval</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">92.98</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">89.46</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">71.35</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><strong>94.27</strong></span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>93.90</u></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value">89.09</span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">-0.37</span></div></td></tr> <tr><td class="benchmark-capability" colspan="8" style="padding:10px 8px;font-weight:700;color:#c2410c;background:rgba(249,115,22,.10);border-top:2px solid #f97316">📊 Overall</td></tr> <tr style="border-bottom:1px solid rgba(128,128,128,.16)"><td class="benchmark-cell" style="padding:8px;font-weight:600"><div class="benchmark-name">Ten-benchmark average</div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">60.57</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">65.60</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">57.29</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value">63.51</span></div></td><td style="padding:8px;text-align:center"><div class="metric-stack"><span class="metric-value"><u>67.86</u></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16)"><div class="metric-stack"><span class="metric-value"><strong>69.04</strong></span></div></td><td style="padding:8px;text-align:center;background:rgba(249,115,22,.16);font-weight:700"><div class="metric-stack"><span class="metric-value">+3.44</span></div></td></tr> </tbody></table> </div>

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 with enable_thinking=true and force_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.

Deployment

Use MLX-LM on an Apple Silicon Mac to run this checkpoint.

Install and select a local checkpoint

bash

pip install "mlx-lm>=0.31.3"
MODEL_PATH="/path/to/NeoHorse-1-9B-MLX"

Set MODEL_PATH to the downloaded MLX directory containing config.json, tokenizer files, chat_template.jinja, and model weights. You can also use TokenRhythm/NeoHorse-1-9B-MLX as the model path to download it automatically from Hugging Face.

Chat locally

bash

mlx_lm.chat --model "$MODEL_PATH"

Start an API server

bash

mlx_lm.server \
  --model "$MODEL_PATH" \
  --host 127.0.0.1 \
  --port 8080

The server exposes an OpenAI-compatible /v1/chat/completions endpoint. In the requests below, default_model refers to the checkpoint selected with --model.

Basic Usage

After the server starts, run this request in another terminal:

bash

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "default_model",
    "messages": [
      {"role": "user", "content": "Write a Python function that returns the first n Fibonacci numbers."}
    ],
    "max_tokens": 2048,
    "stream": false
  }'

The generated reply is returned in choices[0].message.content.

Tool Calling

Pass function definitions in the tools field:

bash

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "default_model",
    "messages": [
      {"role": "user", "content": "Use get_weather to check the current weather in Beijing in celsius."}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get the current weather for a city.",
          "parameters": {
            "type": "object",
            "properties": {
              "city": {"type": "string", "description": "City name."},
              "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
            },
            "required": ["city", "unit"]
          }
        }
      }
    ],
    "max_tokens": 2048,
    "stream": false
  }'

MLX-LM reads the preserved chat template to format tool requests and parse generated calls. When the model chooses to call a tool, the call is returned in choices[0].message.tool_calls. Your application executes the function, appends the assistant message and a role: "tool" result with the matching tool_call_id, then sends the conversation back to the same endpoint for the final answer.

License

NeoHorse-1-9B is released under the Apache License 2.0.

The upstream model is Qwen/Qwen3.5-9B. Its original copyright notice, Copyright 2026 Alibaba Cloud, is retained in the license file. TokenRhythm fine-tuned and repackaged the source checkpoint for text-only inference. This repository provides its MLX BF16 conversion without weight quantization.

Citation

@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 the NeoHorse project repository.