R-4B: Incentivizing General-Purpose Auto-Thinking Capabilities in MLLMs via Bi-Mode Integration

[📚 Arxiv Paper (Coming soon)] [🤗 Hugging Face] [🤖️ ModelScope] [💻 Code]

⭐️ Introduction

In this repo, we present R-4B, a multimodal large language model designed for general-purpose auto-thinking, autonomously switching between step-by-step thinking and direct response generation based on task complexity. This capability enables R-4B to deliver high-quality responses while significantly improving inference efficiency and reducing computational costs.

The development of R-4B follows a two-stage training paradigm: (1) Bi-mode Annealing, which establishes both thinking and non-thinking capabilities for VQA; and (2) Bi-mode Policy Optimization (BPO), which enables the model to adaptively switch between thinking and non-thinking modes based on input demands.

🚀 Key Features

• 🧠 Think Smart, Act Fast: Adaptive & Controllable Thinking! Our model provides three-mode control over the response process.

  • Auto-thinking Mode: Unleash auto-thinking that works across general topics, from simple Q&A to complex scientific analysis. It saves time and computation by thinking only when it matters.
  • Support Manual Control: Explicitly command the model to use its thinking or non-thinking capabilities, enabling you to make your choices for every job.

• 🏆 ** Strong Performance, Open for Everyone!** Our model is now fully open-source. It achieves state-of-the-art performance among models of comparable size.

📢 News

• [2025.08.20] 🚀 vLLM Support is Here! Our R-4B model is now fully compatible with vLLM for high-performance inference.
• [2025.08.18] 🏆 Top Rank Achieved! We are thrilled to announce that R-4B is now ranked #1 among all open-source models on the OpenCompass Multi-modal Reasoning Leaderboard!
• [2025.08.11] 🥇 Another #1! R-4B ranks first under 20B parameters on the OpenCompass Multi-modal Academic Leaderboard!
• [2025.08.05] 🎉 R-4B is Released! Our model is now publicly available. You can download it from Hugging Face.

🔥 Quickstart

Below, we provide simple examples to show how to use R-4B with 🤗 Transformers.

Using 🤗 Transformers to Chat

Users can dynamically control the model's response by selecting one of three modes (auto-thinking, thinking, or non-thinking) with enable_thinking. enable_thinking=auto for auto-thinking mode; enable_thinking=long for thinking mode; enable_thinking=short for non-thinking mode;

import requests
import torch
from transformers import AutoModel, AutoProcessor
from PIL import Image

model_path = "YannQi/R-4B"

model = AutoModel.from_pretrained(
    model_path,
    torch_dtype=torch.float32,
    trust_remote_code=True,
).to("cuda")

processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "http://images.cocodataset.org/val2017/000000039769.jpg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, thinking_mode="auto")

image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)

inputs = processor(images=image, text=text, return_tensors="pt").to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=16384)
output_ids = generated_ids[0][len(inputs.input_ids[0]) :]

# Decode output directly
output_text = processor.decode(
    output_ids,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)

print("Auto-Thinking Output:", output_text)

Using vLLM for fast R-4B deployment and inference.

• We recommend using vLLM for fast R-4B deployment and inference.

Install

The code of R-4B requires the newest vllm now. Please install from local source:

git clone https://github.com/vllm-project/vllm.git
cd vllm
VLLM_USE_PRECOMPILED=1 uv pip install --editable .

Online Serving

The thinking_mode switch is also available in APIs created by vLLM.

• Serve

vllm serve \
    yannqi/R-4B \
    --served-model-name rvl \
    --tensor-parallel-size 8 \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8000 \
    --trust-remote-code

• Openai Chat Completion Client

import base64
from PIL import Image
from openai import OpenAI


# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

# image url
image_messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "http://images.cocodataset.org/val2017/000000039769.jpg"
                },
            },
            {"type": "text", "text": "Describe this image."},
        ],
    },
]

chat_response = client.chat.completions.create(
    model="rvl",
    messages=image_messages,
)
print("Chat response:", chat_response)

📈 Experimental Results

1. R-4B establishes itself with powerful, state-of-the-art perceptual abilities that are competitive with larger models.
2. In evaluation sets that require complex logical reasoning and mathematical problem-solving, such as WeMath, MathVerse, and LogicVista, R-4B displays a strong performance curve. This highlights its advanced adaptive thinking capacity for logical deduction and solving complex quantitative problems.

✒️ Citation

Coming soon!

Acknowledgements

R-4B is developed based on the codebases of the following projects: LLaVA-Next, SigLIP2, Qwen3, Qwen2.5-VL, VLMEvalKit. We sincerely thank these projects for their outstanding work.