TokenRhythm/NeoHorse

NeoHorse is TokenRhythm's family of open-weight models for agent workflows. NeoHorse-1 provides 4B and 9B language models for tool use, coding, and instruction following. NeoHorse-Jev builds on NeoHorse-1-4B to make structured decisions through prefill-only inference.

• [2026-09-24] 🚀 NeoHorse-Jev-4B is here! Our new decision model brings Choice, Noul, and Score to the NeoHorse family. Explore the release.

• [2026-09-09] 📄 Technical report on arXiv! Our technical report is now available, covering the routing harness, agentic post-training, and evaluation of NeoHorse-1.

• [2026-09-08] 📦 GGUF and quantized models on Hugging Face! We release NeoHorse-1-4B-GGUF and NeoHorse-1-9B-GGUF. Both include 16-bit (BF16) weights and smaller 8-bit, 5-bit, and 4-bit quantized versions that use less disk space and memory, making it easier to run NeoHorse on your own hardware. MLX versions for Apple silicon are also available.

• [2026-09-08] 🚀 Now on ModelScope! NeoHorse-1-4B and NeoHorse-1-9B are now available on ModelScope.

• [2026-09-07] 🎉 NeoHorse-1 is here! We release NeoHorse-1-4B and NeoHorse-1-9B under the Apache 2.0 license.

NeoHorse-Jev-4B turns application state into decisions and probabilities. Built on NeoHorse-1-4B, it uses prefill-only inference to select an action with Choice, check a condition with Noul, or assign a rating with Score. Applications define the questions and possible answers, then use the results to route requests, select tools, or control a workflow.

In our evaluation, it scores 77.70 across six text benchmark groups, the highest among the four open-weight decision models with complete results in the comparison. Across Nimble, VitaminC, and MASSIVE, it reaches 83.26% mean accuracy, 11.50 percentage points above NeoHorse-1-4B. See the full comparison and evaluation protocols.

Scores are on a 0–100 scale; higher is better. Bold scores mark the best result and underlined scores the second-best among the listed entries. AVG is the equal-weight mean of the six benchmark groups. -- indicates unavailable results; only models with all six groups are ranked by AVG.

Tetris, Snake, robot manipulation, Mahjong, four-player bomb arena, and autonomous driving — left to right, top to bottom.

NeoHorse-1 is a family of causal language models and an initial prototype on the path toward recursive self-improvement (RSI). The 4B and 9B checkpoints are post-trained from Qwen3.5 for text-based agent harnesses, tool use, coding, and instruction following.

The routing harness assigns tasks to a heterogeneous model pool, records tool interactions and outcomes, estimates capability demand, and feeds capability-level feedback into the next training mixture. Updated models can return to the harness, forming a prototype evaluation–selection–update loop; extending this loop across successive iterations is the next step toward RSI.

• Agentic post-training framework: routing-guided curriculum SFT and routing-guided on-policy distillation turn execution trajectories into training signal while preserving execution and harness context.
• Data quality: exact and near-duplicate removal, evaluation decontamination, structural validation, six-dimensional semantic evaluation, and subscene-level Scene/Goal/Outcome labeling.
• Two release sizes: 4B for a lighter local deployment footprint and 9B for higher capacity on the same text-first serving interface.

NeoHorse-1 checkpoints use a text input/text output interface. NeoHorse-Jev returns structured decisions and probabilities; download its complete model bundle and use the matching runtime or backend adapter. Each model card contains evaluation results and deployment instructions for that model.

The specifications below cover NeoHorse-1-4B and NeoHorse-1-9B. Details for the decision model are in the NeoHorse-Jev documentation.

The tables report NeoHorse-1 results under the ten-benchmark protocol from the technical report. Results are grouped by capability. Higher is better; Δ is NeoHorse minus the same-size Qwen baseline. Bold marks the best result in each benchmark row; ties share the same formatting. For structured decision benchmarks, see the NeoHorse-Jev evaluation.

The 4B comparison includes five representative open-weight baselines. Underline marks the second-best result in each benchmark row.

The 9B comparison includes five representative open-weight baselines from the technical report. Underline marks the second-best result in each benchmark row.

* Nanbeige-4.2-3B LiveCodeBench v6 result is reported from its 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.

For NeoHorse-Jev-4B, follow its deployment guide for vLLM, SGLang, or the native Python runtime, including Choice, Noul, and Score examples.

The examples below serve NeoHorse-1-4B or NeoHorse-1-9B downloaded to local disk. Set MODEL_PATH to the directory containing config.json, tokenizer files, and weights. The request model value is the served alias (neohorse-1-4B or neohorse-1-9B), not the filesystem path.

pip install "sglang==0.5.17"
MODEL_PATH="/path/to/NeoHorse-1-4B"  # or /path/to/NeoHorse-1-9B
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

pip install -U vllm
MODEL_PATH="/path/to/NeoHorse-1-4B"  # or /path/to/NeoHorse-1-9B
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 262,144-token setting is a configured limit; actual capacity depends on GPU memory and serving settings. Reduce it if needed.

The examples below target vLLM at http://127.0.0.1:8000. For SGLang, use http://127.0.0.1:30000. For the 9B checkpoint, use --model neohorse-1-9B and launch the server with --served-model-name neohorse-1-9B.

Install the client dependency:

pip install requests

Run the provided chat example from the repository root:

python examples/chat.py \
  --url http://127.0.0.1:8000 \
  --model neohorse-1-4B

This example sends a predefined prompt to the deployed model and prints the generated response.

Alternatively, send an OpenAI-compatible request directly:

curl http://127.0.0.1:8000/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}'

Run the provided tool-calling example from the repository root:

python examples/tool_call.py \
  --url http://127.0.0.1:8000 \
  --model neohorse-1-4B

This example sends a predefined weather query and prints the tool call generated by the model.

NeoHorse-1 is released under the Apache License 2.0. The upstream models are Qwen3.5-4B and Qwen3.5-9B.

NeoHorse-Jev-4B is also released under Apache License 2.0. Its inference package includes components adapted from Kev; see the NeoHorse-Jev license and acknowledgments for attribution and bundled third-party notices.

@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}
}