Instructions to use Motif-Technologies/Motif-3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Motif-Technologies/Motif-3 with Transformers:

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

pipe = pipeline("text-generation", model="Motif-Technologies/Motif-3", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Motif-Technologies/Motif-3", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Motif-Technologies/Motif-3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Motif-Technologies/Motif-3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Motif-Technologies/Motif-3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Motif-Technologies/Motif-3

SGLang

How to use Motif-Technologies/Motif-3 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 "Motif-Technologies/Motif-3" \
    --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": "Motif-Technologies/Motif-3",
		"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 "Motif-Technologies/Motif-3" \
        --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": "Motif-Technologies/Motif-3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Motif-Technologies/Motif-3 with Docker Model Runner:
```
docker model run hf.co/Motif-Technologies/Motif-3
```

LLAMA CPP suport??

by pawarshardul - opened Aug 10

Discussion

pawarshardul

Aug 10

Any plan for llama cpp support?? This looks very promising !!

eleius

Aug 31

https://github.com/ggml-org/llama.cpp/pull/26298

jhkim55

Sep 3

•

edited Sep 3

Hugging Face CLI

There is now a tested community llama.cpp path for the final Motif-3 core
model. It is not merged into stock llama.cpp yet: the upstream
model-port PR and
GQA-5 Flash Attention PR
are both still open as of September 4.

I built and tested it on one NVIDIA DGX Spark. The released mixed IQ2_XXS GGUF
is 83.56 GiB, keeps all model layers GPU-resident, and measured 316.71 tok/s at
pp512 and 16.49 tok/s at tg128. It was quantized directly from the official
BF16 weights, without a Q5 intermediate.

The runtime builds on Chrono's
(timkhronos) public Motif-3 port and GQA-5
Flash Attention work. My pinned community commit adds the Motif tokenizer
behavior and Unicode regression coverage used for this release. I also
compared the GGUF tokenizer with the official tokenizer across 434,569 token
IDs.

Reproduction notes and exact runtime:
https://github.com/hebo1221/motif3-dgx-spark
GGUF:
https://huggingface.co/jhkim55/Motif-3-Direct-IQ2-XXS-DGX-Spark
Tested runtime commit:
https://github.com/hebo1221/llama.cpp/commit/cc3f13b3f172978d7b3c215780d4cc98bb0e1c80

Two boundaries are worth noting: this GGUF contains the core model rather than
the native MTP head, and the aggressive 2.28 bpw build did not pass my BF16
quality-retention gate. I published the measurements and raw evidence so those
trade-offs are visible rather than implied away.

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