Instructions to use ManiacLabs/Mach-1-Small with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ManiacLabs/Mach-1-Small with Transformers:

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

pipe = pipeline("image-text-to-text", model="ManiacLabs/Mach-1-Small")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("ManiacLabs/Mach-1-Small")
model = AutoModelForMultimodalLM.from_pretrained("ManiacLabs/Mach-1-Small")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ManiacLabs/Mach-1-Small with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ManiacLabs/Mach-1-Small"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ManiacLabs/Mach-1-Small",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/ManiacLabs/Mach-1-Small

SGLang

How to use ManiacLabs/Mach-1-Small 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 "ManiacLabs/Mach-1-Small" \
    --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": "ManiacLabs/Mach-1-Small",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "ManiacLabs/Mach-1-Small" \
        --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": "ManiacLabs/Mach-1-Small",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use ManiacLabs/Mach-1-Small with Docker Model Runner:
```
docker model run hf.co/ManiacLabs/Mach-1-Small
```

Mach-1 Small is part of the Mach-1 family of compressed frontier models, built to run on the machines people actually own.

Mach-1 Small

Qwen3.6-35B-A3B, a 35B-parameter multimodal MoE, compressed to 8.6 GB with the vision tower included. Fits comfortably on 16 GB machines.

~1.4 GB smaller than the smallest public GGUF of this model (which ships without vision) at comparable answer fidelity
116.7 tok/s on an M5 Max with the Mach engine (162.3 tok/s with speculative decoding), ~7 GB resident for text sessions, 6-second warm load

Run it

Mach inference engine: native format, fastest path.
Any transformers stack: materialize a standard bf16 checkpoint (~70 GB) with the included decoder, then load with AutoModelForImageTextToText:

pip install torch numpy safetensors zstandard huggingface_hub
python reconstruct.py --repo ManiacLabs/Mach-1-Small --out ./mach1-small

Quality

Measured against the full-precision model on a fixed 2000-question MMLU stratum, on a model whose weights decode 100% from the files in this repo:

	Mach-1 Small	smallest public GGUF*
total size (GB, with vision)	8.62	10.05*
answer agreement vs teacher	88.6%	89.5%
MMLU accuracy (teacher: 82.6%)	79.4%	79.6%

On a 4x larger (8000-question) stratum: 88.8% agreement, 80.1% accuracy (teacher 83.2% there).

Coding (Aider Polyglot), 225 exercises, pass-2, identical serving path for every arm:

	teacher bf16 (~70 GB)	Mach-1 Small	Unsloth UD-IQ1_M (10.05 GB*)
exercises solved	39	28	18

*Unsloth UD-IQ1_M (9.2 GB, no vision tower; +0.87 GB bf16-vision parity added for a fair total), identical harness and teacher.

Notes

All quality numbers were measured on weights decoded from the files in this repo, and the decoder that proves it (decode.py, reconstruct.py) ships here, pure numpy.
The evaluation harness is text-only. The vision tower ships bf16 (unquantized); treat multimodal quality as inherited from the base model, not validated by us.
The MTP speculative-decoding head is not shipped (standard generation never runs it). Reconstructed checkpoints show it as missing keys; that is expected.
Everything in this repo is Apache-2.0, like the base model.

Downloads last month: 126

Model tree for ManiacLabs/Mach-1-Small

Base model

Qwen/Qwen3.6-35B-A3B

Finetuned

(166)

this model