Last update: 10 Dec. 2025

Introduction

This is a reasoning enhanced version of Motif-2-12.7B-Instruct. Detailed information will be released later.

Evaluation

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 64K tokens.

# run vllm api server
VLLM_ATTENTION_BACKEND=DIFFERENTIAL_FLASH_ATTN \
vllm serve Motif-Technologies/Motif-2-12.7B-Reasoning \
    --trust-remote-code \
    --max-model-len 65536 \
    --tensor-parallel-size 8

# 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
  }'

How to use advanced vllm options

For maximum performance, we highly recommend using the options below.
--compilation_config '{"full_cuda_graph": true}' : Activates cuda full graph capture
--rope-scaling '{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":65536}': Apply yarn to support 128K context length
--enable-auto-tool-choice --tool-call-parser hermes : Enables tool calling
--logits-processors logit_:WrappedPerReqLogitsProcessor: Enables a ratio-based thinking budget and repetition-based auto-stop. The model is guided to think for (model_max_len - input_prompt_len) * VLLM_THINK_BUDGET_RATIO tokens, using the rest of the context window to generate the response
--reasoning-parser deepseek_r1 : Parses reasoning outputs

pip install -U "huggingface_hub[cli]"
hf download Motif-Technologies/Motif-2-12.7B-Reasoning \
  --include "logit_processors/*" \
  --local-dir ./

export PYTHONPATH="$PWD/logit_processors"
VLLM_ATTENTION_BACKEND=DIFFERENTIAL_FLASH_ATTN \
VLLM_THINK_BUDGET_RATIO=0.95 \ 
vllm serve Motif-Technologies/Motif-2-12.7B-Reasoning \
    --trust-remote-code \
    --compilation_config '{"full_cuda_graph": true}' \
    --rope-scaling '{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":65536}' \
    --max-model-len 131072 \
    --tensor-parallel-size 8 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes \
    --logits-processors logit_:WrappedPerReqLogitsProcessor \
    --reasoning-parser deepseek_r1