Instructions to use Motif-Technologies/Motif-2-12.7B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Motif-Technologies/Motif-2-12.7B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="Motif-Technologies/Motif-2-12.7B-Instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Motif-Technologies/Motif-2-12.7B-Instruct", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Motif-Technologies/Motif-2-12.7B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Motif-Technologies/Motif-2-12.7B-Instruct"
# 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-2-12.7B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Motif-Technologies/Motif-2-12.7B-Instruct

SGLang

How to use Motif-Technologies/Motif-2-12.7B-Instruct 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-2-12.7B-Instruct" \
    --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-2-12.7B-Instruct",
		"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-2-12.7B-Instruct" \
        --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-2-12.7B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Motif-Technologies/Motif-2-12.7B-Instruct with Docker Model Runner:
```
docker model run hf.co/Motif-Technologies/Motif-2-12.7B-Instruct
```

Last update: 12 Nov. 2025

Introduction

We are pleased to announce Motif-2-12.7B-Instruct, a 12.7-billion-parameter language model. This model is an supervised fine-tuning (SFT) variant of our base model: https://huggingface.co/Motif-Technologies/Motif-2-12.7B-Base. Detailed information is found in the technical report: https://arxiv.org/abs/2511.07464.

One can chat directly with Motif-2-12.7B-Instruct at https://chat.motiftech.io.

Evaluation

The results of Qwen3 and Gemma 3 are sourced directly from their technical reports.

Benchmark	Evaluation setting	Motif-2-12.7B	Qwen2.5-72B	Qwen3-14B	Qwen3-14B	Qwen3-32B	Qwen3-32B	Qwen3-30B-A3B	Qwen3-30B-A3B	Gemma-3-12B	Gemma-3-27B
		Instruct	Instruct	Non-thinking	Thinking	Non-thinking	Thinking	Non-thinking	Thinking	Instruct	Instruct
MMLU	0-shot	86.11	-	-	-	-	-	-	-	71.9	76.9
MMLU-Redux	-	90.02	86.8	82	88.6	85.7	90.9	84.1	89.5	-	-
BBH	0-shot	85.78	-	-	-	-	-	-	-	85.7	87.6
GPQA-Diamond	0-shot, CoT	63.6	49	54.8	64	54.6	68.4	54.8	65.8	40.9	42.4
GSM8K	0-shot, CoT	96.13	-	-	-	-	-	-	-	94.4	95.9
MATH	0-shot	97	-	-	-	-	-	-	-	83.8	89
MBPP	3-shot	91	-	-	-	-	-	-	-	73	74.4
LiveBench 2024-11-25	-	33.8	51.4	59.6	71.3	59.8	74.9	59.4	74.3	-	-
IFEval	strict prompt	75.78	84.1	84.8	85.4	83.2	85	83.7	86.5	-	-
IFEval	0-shot	76.52	-	-	-	-	-	-	-	88.9	90.4
MATH-500	-	96.8	83.6	90	96.8	88.6	97.2	89.8	98	-	-
AIME24	-	72.3	18.9	31.7	79.3	31	81.4	32.8	80.4	-	-
AIME25	-	63.6	15	23.3	70.4	20.2	72.9	21.6	70.9	-	-
ZebraLogic	-	69.5	26.6	33	88.5	29.2	88.8	33.2	89.5	-	-
BFCL v3	-	55.34	63.4	61.5	70.4	63	70.3	58.6	69.1	-	-
LiveCodeBench v5 (2024.10 - 2025.2)	-	50.03	30.7	29	63.5	31.3	65.7	29.8	62.6	-	-
LiveCodeBench v5	0-shot, CoT	61.66	-	-	-	-	-	-	-	32	39
HumanEval	0-shot	93.2	-	-	-	-	-	-	-	85.4	87.8

Averages and improvements of the corresponding benchmark scores:

v.s. Gemma 3

	Motif-2-12.7B	Gemma-3-12B	Gemma-3-27B
	Instruct	Instruct	Instruct
Average	83.44	72.89	75.93
Improvement		+14.48%	+9.89%

v.s. Qwen3

	Motif-2-12.7B	Qwen2.5-72B	Qwen3-14B	Qwen3-14B	Qwen3-32B	Qwen3-32B	Qwen3-30B-A3B	Qwen3-30B-A3B
	Instruct	Instruct	Non-thinking	Thinking	Non-thinking	Thinking	Non-thinking	Thinking
Average	67.08	50.95	54.97	77.82	54.66	79.55	54.78	78.66
Improvement		+31.65%	+22.02%	-13.80%	+22.72%	-15.68%	+22.45%	-14.73%

How to use in transformers

To use this model, install huggingface kernels.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Motif-Technologies/Motif-2-12.7B-Instruct",
    trust_remote_code = True,
    _attn_implementation = "flash_attention_2",
    dtype = torch.bfloat16 # currently supports bf16 only, for efficiency
).cuda()

tokenizer = AutoTokenizer.from_pretrained(
    "Motif-Technologies/Motif-2-12.7B-Instruct",
    trust_remote_code = True,
)

query = "What is the capital city of South Korea?"
input_ids = tokenizer.apply_chat_template(
    [
        {'role': 'system', 'content': 'you are an helpful assistant'},
        {'role': 'user', 'content': query},
    ],
    add_generation_prompt = True,
    enable_thinking = False, # or True
    return_tensors='pt',
).cuda()

output = model.generate(input_ids, max_new_tokens=1024, pad_token_id=tokenizer.eos_token_id)
output = tokenizer.decode(output[0, input_ids.shape[-1]:], skip_special_tokens = False)
print(output)

outputs

# with enable_thinking=True, the model is FORCED to think.
Okay, the user is asking for the capital city of South Korea. Let me think. I know that South Korea's capital is Seoul. But wait, I should double-check to make sure I'm not mixing it up with other countries. For example, North Korea's capital is Pyongyang. So yes, South Korea's capital is definitely Seoul. I should just provide that as the answer.
</think>
The capital city of South Korea is **Seoul**.
<|endofturn|><|endoftext|>

# with enable_thinking=False, the model chooses to think or not. in this example, thinking is not worth it. 
The capital city of South Korea is Seoul.
<|endofturn|><|endoftext|>

How to use in vllm

The PR adding support for the Motif model in the official vLLM package is currently under review.
In the meantime, to use our model with vLLM, please use the following container image.
Our model supports a sequence length of up to 32K tokens.

# run vllm api server
VLLM_ATTENTION_BACKEND="DIFFERENTIAL_FLASH_ATTN" vllm serve Motif-Technologies/Motif-2-12.7B-Instruct --trust-remote-code --data-parallel-size <gpu_count>

# sending requests with curl
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "What is the capital city of South Korea?"}
    ],
    "temperature": 0.6,
    "skip_special_tokens": false,
    "chat_template_kwargs": {
        "enable_thinking": true
    }
  }'