Skywork-R1V2-38B-AWQ

📖 R1V2 Report | 💻 GitHub | 🌐 ModelScope

Evaluation

Comprehensive performance comparison across text and multimodal reasoning benchmarks.

Model	MMMU	MathVista	MathVision	Olympiad Bench	AIME 24	LiveCode bench	Live Bench	IFEVAL
Proprietary Models
Claude-3.5-Sonnet	70.4	67.7	-	-	-	-	-	-
Gemini-2-Flash	70.7	73.1	41.3	-	-	-	-	-
Kimi-k1.5-longcot	70.0	74.9	53.3	-	-	-	-	-
OpenAI-o1	-	-	-	-	74.3	63.4	72.2	-
OpenAI-o4-mini	81.6	84.3	58.0	-	93.4	74.6	78.1	-
Open-Source Models
Skywork-R1V1	68.0	67.0	-	-	72.0	57.2	54.6	72.5
DeepseekR1-671B	-	-	-	-	79.8	65.9	71.6	83.3
InternVL3-38B	70.1	75.1	34.2	-	-	-	-	-
Qwen2.5-VL-72B	70.2	74.8	38.1	40.4	-	-	-	-
QvQ-Preview-72B	70.3	71.4	35.9	33.2	-	-	-	-
Skywork-R1V2	73.6	74.0	49.0	62.6	78.9	63.6	73.2	82.9
Skywork-R1V2-AWQ	64.4	64.8	42.9	54.8	77.3	55.7	64.1	72.5

Usage

You can use the quantized model with different inference frameworks:

Using VLLM

Python API

import os
from vllm import LLM, SamplingParams
from vllm.entrypoints.chat_utils import load_chat_template
model_name = "Skywork/Skywork-R1V2-38B-AWQ"  # or local path
llm = LLM(model_name, 
          dtype='float16', 
          quantization="awq", 
          gpu_memory_utilization=0.9,
          max_model_len=4096,
          trust_remote_code=True,
         )
# Add your inference code here

OpenAI-compatible API Server

MODEL_ID="Skywork/Skywork-R1V2-38B-AWQ"  # or local path
CUDA_VISIBLE_DEVICES=0 \
    python -m vllm.entrypoints.openai.api_server \
    --model $MODEL_ID \
    --dtype float16 \
    --quantization awq \
    --port 23334 \
    --max-model-len 12000 \
    --gpu-memory-utilization 0.9 \
    --trust-remote-code

Using LMDeploy

import os
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
from lmdeploy.vl import load_image
model_path = "Skywork/Skywork-R1V2-38B-AWQ"  # or local path
engine_config = TurbomindEngineConfig(cache_max_entry_count=0.75) 
chat_template_config = ChatTemplateConfig(model_name=model_path)
pipe = pipeline(model_path, 
                backend_config=engine_config, 
                chat_template_config=chat_template_config,
               )
# Example: Multimodal inference
image = load_image('table.jpg')
response = pipe(('Describe this image?', image))
print(response.text)

Hardware Requirements

The AWQ quantization reduces the memory footprint compared to the original FP16 model. We recommend:

• At least one GPU with 30GB+ VRAM for inference
• For optimal performance with longer contexts, 40GB+ VRAM is recommended

Citation

If you use this model in your research, please cite:

@misc{peng2025skyworkr1vpioneeringmultimodal,
      title={Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought}, 
      author={Yi Peng and Chris and Xiaokun Wang and Yichen Wei and Jiangbo Pei and Weijie Qiu and Ai Jian and Yunzhuo Hao and Jiachun Pan and Tianyidan Xie and Li Ge and Rongxian Zhuang and Xuchen Song and Yang Liu and Yahui Zhou},
      year={2025},
      eprint={2504.05599},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.05599}, 
}

@misc{chris2025skyworkr1v2multimodalhybrid,
      title={Skywork R1V2: Multimodal Hybrid Reinforcement Learning for Reasoning}, 
      author={Chris and Yichen Wei and Yi Peng and Xiaokun Wang and Weijie Qiu and Wei Shen and Tianyidan Xie and Jiangbo Pei and Jianhao Zhang and Yunzhuo Hao and Xuchen Song and Yang Liu and Yahui Zhou},
      year={2025},
      eprint={2504.16656},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.16656}, 
}