Instructions to use TokenRhythm/NeoHorse-1-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TokenRhythm/NeoHorse-1-4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="TokenRhythm/NeoHorse-1-4B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("TokenRhythm/NeoHorse-1-4B")
model = AutoModelForCausalLM.from_pretrained("TokenRhythm/NeoHorse-1-4B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use TokenRhythm/NeoHorse-1-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TokenRhythm/NeoHorse-1-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TokenRhythm/NeoHorse-1-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/TokenRhythm/NeoHorse-1-4B

SGLang

How to use TokenRhythm/NeoHorse-1-4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "TokenRhythm/NeoHorse-1-4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TokenRhythm/NeoHorse-1-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "TokenRhythm/NeoHorse-1-4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TokenRhythm/NeoHorse-1-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use TokenRhythm/NeoHorse-1-4B with Docker Model Runner:
```
docker model run hf.co/TokenRhythm/NeoHorse-1-4B
```

Solid work

by LonelyDance2009 - opened Sep 11

Discussion

LonelyDance2009

Sep 11

Solid work — the routing harness framing makes the "self-improvement" loop concrete rather than hand-wavy, and +5.93 at 4B is a real jump. Two small things: the chat template still renders the <|vision_start|><|image_pad|><|vision_end|> path even though only LM weights are included, so image input may fail silently rather than erroring — worth a note in the card. And VitaBench using DeepSeek-V4-Flash as both simulator and judge, on top of being single-run with a +10.50 delta, is the number I'd most want to see cross-checked. Thanks for the Apache-2.0 release.

JarvisPei

TokenRhythm org Sep 11

Solid work — the routing harness framing makes the "self-improvement" loop concrete rather than hand-wavy, and +5.93 at 4B is a real jump. Two small things: the chat template still renders the <|vision_start|><|image_pad|><|vision_end|> path even though only LM weights are included, so image input may fail silently rather than erroring — worth a note in the card. And VitaBench using DeepSeek-V4-Flash as both simulator and judge, on top of being single-run with a +10.50 delta, is the number I'd most want to see cross-checked. Thanks for the Apache-2.0 release.

Thanks for the thoughtful feedback! This release is text-only; the image/video template branches are inherited and don’t imply vision support. We’ll clarify that in the model card.
Agreed on VitaBench: the +10.50 is a single-run result under the disclosed simulator/judge setup. Repeated runs and an independent judge would be useful cross-checks. Thanks for flagging both!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment