--- license: apache-2.0 library_name: mlx pipeline_tag: text-generation base_model: TokenRhythm/NeoHorse-1-4B tags: - agentic - tool-use - coding - reasoning - instruction-following - mlx - bf16 --- ## MLX local inference This is the **unquantized BF16 MLX** version of [NeoHorse-1-4B](https://huggingface.co/TokenRhythm/NeoHorse-1-4B) 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-1-4B-MLX ``` The model downloads automatically from Hugging Face. The original chat template is preserved. See [Deployment](#deployment) for local checkpoints, the chat API, and tool calling.

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](https://huggingface.co/Qwen/Qwen3.5-4B) 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:** 64.87 macro average across ten benchmarks versus 58.94 for Qwen3.5-4B (**+5.93**). ## Model Details

Property	Value
Model family	NeoHorse Agent-Native Causal Language Model
Parameters	Approximately 4B
Base model	Qwen3.5-4B
Post-training	Routing-guided agentic post-training
Interface	Text input and text output
Context length	262,144 natively and extensible up to 1,010,000 tokens.
Weight format / precision	MLX Safetensors / BF16 (unquantized)

## 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.

Benchmark	Qwen3.5-4B	Gemma-4-E4B-it	Nanbeige-4.2-3B	Agents-A1-4B	Spark-X2.5-4B	NeoHorse-1-4B	Δ vs Qwen3.5-4B
🤖 Agentic
QwenClawBench	38.47	22.98	40.66	43.16	43.52	44.68	+6.21
WorkBuddy Bench	24.62	11.65	21.03	33.37	26.47	34.41	+9.79
PinchBench	71.19	47.60	66.78	75.07	62.37	77.33	+6.14
VitaBench	21.50	5.00	31.50	39.25	37.00	32.00	+10.50
BFCL v4	61.02	47.18	67.28	46.60	63.71	61.79	+0.77
tau2-Bench	84.29	43.60	85.08	81.00	77.72	88.46	+4.17
💻 Coding
HumanEval	87.20	84.76	98.78	92.68	92.07	96.95	+9.75
LiveCodeBench v6	53.71	52.00	72.50^*	56.57	54.86	59.43	+5.72
📚 Instruction Following
IFBench	60.33	40.00	55.00	63.33	73.33	65.33	+5.00
IFEval	87.06	74.68	84.47	83.55	91.13	88.35	+1.29
📊 Overall
Ten-benchmark average	58.94	42.95	62.31	61.46	62.22	64.87	+5.93

`*` 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 Use [MLX-LM](https://github.com/ml-explore/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-1-4B-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-1-4B-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-1-4B is released under the **Apache License 2.0**. The upstream model is [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B). 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](https://github.com/TokenRhythm/NeoHorse).