NeoHorse-1-4B

Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness.

Technical Report

NeoHorse-1-4B is a 4B causal language model and an initial prototype on the path toward recursive self-improvement (RSI). It is post-trained from Qwen3.5-4B for text-based agent harnesses, tool use, coding, and instruction following.

Derived from Qwen/Qwen3.5-4B and fine-tuned by TokenRhythm. This release contains language-model weights only and 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.

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: 64.87 macro average across ten benchmarks versus 58.94 for Qwen3.5-4B (+5.93).

Model Details

Evaluation

The 4B track compares NeoHorse-1-4B with five representative open-weight models. Results are grouped by capability in the table below. Higher is better; Δ is NeoHorse-1-4B minus Qwen3.5-4B. Bold marks the best available result; underlining marks the second-best.

* Nanbeige-4.2-3B LiveCodeBench v6 result is reported in the corresponding model's official blog post or technical report.

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

The examples below are for self-hosted deployment from a downloaded local checkpoint.

Local checkpoint path

The examples below assume the checkpoint has already been downloaded to local disk. Set MODEL_PATH to the directory containing config.json, tokenizer files, and model weights.

MODEL_PATH="/path/to/NeoHorse-1-4B"

The OpenAI-compatible requests below use the server's --served-model-name (for example, neohorse-1-4b), not the filesystem path.

SGLang

The technical report uses SGLang v0.5.17.

pip install "sglang==0.5.17"
MODEL_PATH="/path/to/NeoHorse-1-4B"
python3 -m sglang.launch_server \
  --model-path "$MODEL_PATH" \
  --served-model-name neohorse-1-4b \
  --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-1-4b","messages":[{"role":"user","content":"Write a Python function that returns the first n Fibonacci numbers."}],"max_tokens":512}'

vLLM

pip install -U vllm
MODEL_PATH="/path/to/NeoHorse-1-4B"
vllm serve "$MODEL_PATH" \
  --served-model-name neohorse-1-4b \
  --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/completions endpoint. Send a request after the server starts:

curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"neohorse-1-4b","messages":[{"role":"user","content":"Write a Python function that returns the first n Fibonacci numbers."}],"max_tokens":512}'

The examples use the configured 262,144-token context limit. Actual capacity depends on GPU memory and serving settings; reduce the context limit if needed.

License

NeoHorse-1-4B is released under the Apache License 2.0.

The upstream model is Qwen/Qwen3.5-4B. 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.

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.