OneThink
/

OneThinker-8B

@@ -11,15 +11,19 @@ library_name: transformers
 This repository contains the model presented in: [OneThinker: All-in-one Reasoning Model for Image and Video](https://huggingface.co/papers/2512.03043)
 **Project Page**: https://github.com/tulerfeng/OneThinker
 **Code**: https://github.com/tulerfeng/OneThinker
 ## 👀 About OneThinker
 <div align="center">
-  <img src="https://huggingface.co/datasets/OneThink/OneThinker-8B/resolve/main/assets/teaser.png" alt="Descriptive alt text" width="95%">
-</div>
 We introduce **OneThinker**, an all-in-one multimodal reasoning generalist that is **capable of thinking across a wide range of fundamental visual tasks within a single model**.
 We construct the large-scale **OneThinker-600k** multi-task training corpus and build **OneThinker-SFT-340k** with high-quality CoT annotations for cold-start SFT. Moreover, we propose **EMA-GRPO**, a new RL method that **balances heterogeneous reward signals across diverse visual tasks**, via simply tracking task-wise moving averages of reward std.
@@ -28,32 +32,13 @@ OneThinker demonstrates **strong performance on 31 benchmarks across 10 fundamen
 All code, models, and data are fully released.
-## 🔥 News
-- [2025/12/03] We release the code, model, data of OneThinker
-## 📍 Features
-+ Support Qwen3-VL Training
-+ Support Image-Video mixed training
-+ Support reward types in diverse visual tasks
-+ Provide full pipeline (dataset, SFT training, RL training, evaluation, etc)
-## 🔍 Dataset
-Our dataset covers both image and video modalities and spans a series of fundamental visual reasoning tasks, including rule-based QA, open-ended QA, captioning, spatial grounding, temporal grounding, spatio-temporal grounding, tracking, and segmentation
-<div align="center">
-  <img src="https://huggingface.co/datasets/OneThink/OneThinker-8B/resolve/main/assets/dataset.png" alt="Descriptive alt text" width="90%">
-</div>
-To enable effective SFT initialization for reasoning, we leverage a strong proprietary model, Seed1.5-VL  to produce CoT annotations.
 ## 🏆 Performance
 Our model obtains significant performance gains after training based on Qwen3-VL-Instruct-8B across diverse visual tasks. For examle, OneThinker-8B reaches 70.6% accuracy on MMMU, 64.3% on MathVerse, 66.2% on VideoMMMU, 93.7 on Refcoco-testA, 54.9 J&F on ReasonVOS.
 <div align="center">
-  <img src="https://huggingface.co/datasets/OneThink/OneThinker-8B/resolve/main/assets/performance.png" alt="Descriptive alt text" width="90%">
 </div>
 Besides, we also observe beneficial cross-task and cross-modality knowledge transfer, along with promising preliminary zero-shot generalization under unified training. This highlights the  effectiveness and generalization ability of our unified training framework across diverse visual tasks.
@@ -66,7 +51,7 @@ Some reasoning examples are as follows.
 <summary>Demo 1 (QA)</summary>
 <div align="center">
-  <img src="https://huggingface.co/datasets/OneThink/OneThinker-8B/resolve/main/assets/math.png" width="36%">
 </div>
 **Question:**
@@ -96,7 +81,7 @@ AB is diameter, so ∠ACB = 90°, ∠ADB = 90°. In △ABD, ∠BAD = 90° - 50°
 <summary>Demo 2 (Tracking)</summary>
 <div align="center">
-  <img src="https://huggingface.co/datasets/OneThink/OneThinker-8B/resolve/main/assets/got_car.gif" width="60%">
 </div>
 **Question:**
@@ -116,7 +101,7 @@ Given the bounding box [537,403,768,703] of the target object in the first frame
 <summary>Demo 3 (Segmentation)</summary>
 <div align="center">
-  <img src="https://huggingface.co/datasets/OneThink/OneThinker-8B/resolve/main/assets/lalaland.gif" width="60%">
 </div>
 **Question:**
@@ -132,156 +117,6 @@ Pleae segment the shorter one.
 </details>
-## 🚀 Quick Start (Inference)
-Here's a simple example to run inference with the OneThinker model using the Hugging Face `transformers` library:
-```python
-import numpy as np
-import torch
-from PIL import Image
-from transformers import AutoProcessor, AutoModelForCausalLM
-# Load model and processor
-model_id = "OneThink/OneThinker-8B" # This model
-processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True).eval()
-# Example 1: Image QA
-print("--- Image QA Demo ---")
-# Replace with a valid local path to your image, e.g., downloaded from the GitHub repo or assets
-image_path = "./assets/math.png"
-try:
-    image = Image.open(image_path).convert("RGB")
-except FileNotFoundError:
-    print(f"Warning: Image '{image_path}' not found. Using a placeholder image.")
-    image = Image.new('RGB', (500, 500), color = 'blue') # Placeholder if image not found
-prompt_qa = "As shown in the figure, AB is the diameter of ⊙O, and points C and D are on ⊙O. If ∠ABD = 50.0, then the degree of ∠BCD is () Choices: (A) 30° (B) 35° (C) 40° (D) 45°"
-messages_qa = [
-    {"role": "user", "content": [
-        {"type": "image", "content": image},
-        {"type": "text", "content": prompt_qa}
-    ]}
-]
-input_ids = processor.apply_chat_template(messages_qa, tokenize=True, add_generation_prompt=True, return_tensors="pt", max_new_tokens=2048)
-inputs = processor(messages_qa, return_tensors="pt")
-inputs = {k: v.to(model.device) for k, v in inputs.items()}
-with torch.no_grad():
-    output_ids = model.generate(**inputs, max_new_tokens=2048)
-    output_text = processor.batch_decode(output_ids[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]
-print(f"Question: {prompt_qa}
-Answer: {output_text}
-")
-# Example 2: Video Tracking (Conceptual/Simplified)
-print("--- Video Tracking Demo (Conceptual/Simplified) ---")
-# For full video processing, refer to the scripts in the original GitHub repository.
-# This is a simplified example showing how to pass video frames.
-# In a real scenario, you would load a sequence of actual video frames:
-# video_frames_list = [Image.open(f"path/to/video_frame_{i:04d}.jpg").convert("RGB") for i in range(num_frames)]
-# For demonstration, we'll use a list of placeholder images:
-video_frames_list = [Image.new('RGB', (224, 224), color = 'red') for _ in range(5)] # 5 placeholder frames
-prompt_tracking = "Given the bounding box [537,403,768,703] of the target object in the first frame, track this object in each frame."
-messages_tracking = [
-    {"role": "user", "content": [
-        {"type": "video", "content": video_frames_list}, # Pass video frames directly
-        {"type": "text", "content": prompt_tracking}
-    ]}
-]
-inputs_tracking = processor(messages_tracking, return_tensors="pt")
-inputs_tracking = {k: v.to(model.device) for k, v in inputs_tracking.items()}
-with torch.no_grad():
-    output_ids_tracking = model.generate(**inputs_tracking, max_new_tokens=2048)
-    output_text_tracking = processor.batch_decode(output_ids_tracking[:, inputs_tracking["input_ids"].shape[1]:], skip_special_tokens=True)[0]
-print(f"Question: {prompt_tracking}
-Answer: {output_text_tracking}
-")
-```
-## 📐 Set up
-```bash
-git clone https://github.com/tulerfeng/OneThinker
-cd OneThinker
-# build SFT environment
-conda create -n llamafactory python=3.11
-conda activate llamafactory
-cd LLaMA-Factory
-pip install -e ".[torch,metrics]" --no-build-isolation
-# build RL environment
-conda create -n easyr1 python=3.11
-conda activate easyr1
-cd EasyR1
-pip install -e .
-```
-For more details for the SFT and RL environment installation, please refer to [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory),  [EasyR1](https://github.com/hiyouga/EasyR1)
-Then, download the training datasets [[🤗 OneThinker-train-data](https://huggingface.co/datasets/OneThink/OneThinker-train-data)] and unzip all the data.
-The `onethinker_rl_train.json` file is for RL training while `onethinker_sft_image.json` and `onethinker_sft_video.json` is for SFT cold start. The json files end with `_unsampled` are unsampled full set.
-## 🚀 Training
-For SFT and RL training, a minimum of 8 × 80GB GPUs is required; alternatively, you may reduce the number of frames or the input resolution.
-We first perform SFT cold start.
-```bash
-bash ./LLaMA-Factory/local_scripts/run_onethinker_sft.sh
-```
-If you want to skip the SFT process, we also provide our SFT model at [🤗[OneThinker-SFT-model](https://huggingface.co/OneThink/OneThinker-SFT-Qwen3-8B)]
-Then, we perform RL training  as follows
-```bash
-bash ./EasyR1/local_scripts/run_onethinker_rl.sh
-```
-For setting Ray in multi-node training, please refer to [EasyR1](https://github.com/hiyouga/EasyR1), or you may use single-node training by setting `NNODES=1`. Performing RL training for about 200 steps can already yield strong performance.
-If you want to use model-based rewards for open-ended problem, please use vllm to lanuch [POLAR-7B](https://github.com/InternLM/POLAR) and revised the setting in `/EasyR1/verl/reward_function/onethinker_reward.py`
-## 🔮 Inference & Evaluation
-For the majority of tasks and benchmarks, we recommend using our provided json files and scripts for easier evaluation.
-The json files can be downloaded at: [🤗 [OneThinker-eval](https://huggingface.co/datasets/OneThink/OneThinker-eval)]
-Download the trained model [[🤗 OneThinker-8B-model](https://huggingface.co/OneThink/OneThinker-8B)]
-Conduct evaluation on all benchmarks using the following scripts
-```bash
-bash ./Evaluation/Eval/eval_bench_all.sh
-```
-If you want to perform evaluation on segmentation tasks, please download and install [sam2](https://github.com/facebookresearch/sam2) and revise the related path in `/Evaluation/Eval/seg_post_sam2.py`
-For image QA and part of video QA, we use  [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) for evaluation, please install corresponding environment and run:
-```bash
-bash ./Evaluation/VLMEvalKit/local_scripts/eval_vlmevalkit.sh
-```
-For infernce on a single example, you may refer to:
-```bash
-python ./Evaluation/inference_single/inference.py
-```
-## Acknowledgements
-We sincerely appreciate the contributions of the open-source community. The related projects are as follows: [Video-R1](https://github.com/tulerfeng/Video-R1), [DeepSeek-R1](https://github.com/deepseek-ai/DeepSeek-R1), [EasyR1](https://github.com/hiyouga/EasyR1), [verl](https://github.com/volcengine/verl),  [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory), [VLMEvalKit](https://github.com/open-compass/VLMEvalKit)
 ## Citations

 This repository contains the model presented in: [OneThinker: All-in-one Reasoning Model for Image and Video](https://huggingface.co/papers/2512.03043)
+For inference, please refer to:
 **Project Page**: https://github.com/tulerfeng/OneThinker
 **Code**: https://github.com/tulerfeng/OneThinker
 ## 👀 About OneThinker
 <div align="center">
+  <img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/teaser.png?raw=true" alt="OneThinker teaser" width="95%">
+</div>
 We introduce **OneThinker**, an all-in-one multimodal reasoning generalist that is **capable of thinking across a wide range of fundamental visual tasks within a single model**.
 We construct the large-scale **OneThinker-600k** multi-task training corpus and build **OneThinker-SFT-340k** with high-quality CoT annotations for cold-start SFT. Moreover, we propose **EMA-GRPO**, a new RL method that **balances heterogeneous reward signals across diverse visual tasks**, via simply tracking task-wise moving averages of reward std.
 All code, models, and data are fully released.
 ## 🏆 Performance
 Our model obtains significant performance gains after training based on Qwen3-VL-Instruct-8B across diverse visual tasks. For examle, OneThinker-8B reaches 70.6% accuracy on MMMU, 64.3% on MathVerse, 66.2% on VideoMMMU, 93.7 on Refcoco-testA, 54.9 J&F on ReasonVOS.
 <div align="center">
+  <img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/performance.png?raw=true" alt="Descriptive alt text" width="90%">
 </div>
 Besides, we also observe beneficial cross-task and cross-modality knowledge transfer, along with promising preliminary zero-shot generalization under unified training. This highlights the  effectiveness and generalization ability of our unified training framework across diverse visual tasks.
 <summary>Demo 1 (QA)</summary>
 <div align="center">
+  <img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/math.png?raw=true" width="36%">
 </div>
 **Question:**
 <summary>Demo 2 (Tracking)</summary>
 <div align="center">
+  <img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/got_car.gif?raw=true" width="60%">
 </div>
 **Question:**
 <summary>Demo 3 (Segmentation)</summary>
 <div align="center">
+  <img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/lalaland.gif?raw=true" width="60%">
 </div>
 **Question:**
 </details>
 ## Citations