Mini-o3 - A visual reasoning model jointly developed by ByteDance and the University of Hong Kong

What is Mini-o3?

Mini-o3 is an open-source model jointly developed by ByteDance and the University of Hong Kong, designed specifically to solve complex visual search problems. Through reinforcement learning and image-based tools, the model can perform deep multi-turn inference, with the number of inference rounds scalable to dozens. Mini-o3 significantly improves the model's inference ability and interaction rounds by constructing challenging datasets, adopting iterative data collection processes, and proposing an over-turn occlusion strategy. The model achieves state-of-the-art performance on multiple visual search benchmarks. All code, the model, and the dataset are open-source, facilitating reproduction and further research.

Main functions of Mini-o3

• Multi-turn interactive reasoningIt can perform deep, multi-round reasoning, with the number of interaction rounds expandable to dozens, solving complex visual search problems through gradual exploration and trial and error.
• Diverse Reasoning PatternsIt supports multiple inference modes, including depth-first search, trial and error, and goal maintenance.
• Challenging visual searchEven in high-resolution images, where the target is small and there are many interfering objects, the target can still be accurately located and identified.
• Excellent performanceIt achieves state-of-the-art results in multiple visual search benchmarks (such as VisualProbe, V* Bench, HR-Bench, MME-Realworld, etc.), demonstrating powerful visual reasoning capabilities.
• open sourceAll code, models, and datasets are open source, making it easy for researchers to reproduce and conduct further research, thus promoting the development of related technologies.

Mini-o3's technical principles

• Cold-start Supervised Fine-tuning (SFT): Using a small number of handcrafted examples, we generate high-quality, diverse, multi-turn interaction trajectories with a context-learning visual language model (VLM).
• Reinforcement Learning (RL)Based on the over-turn masking strategy, it avoids being penalized for exceeding the limit of interaction rounds and can naturally expand to dozens of interaction rounds during testing.
• Reduce image pixel budget (Lower Down Max Pixels)By reducing the maximum number of pixels per image, the number of rounds allowed in a single interaction is increased, thereby improving the ability to solve long-cycle problems.
• Challenging Dataset (Visual Probe Dataset): Construct a dataset containing thousands of visual search questions designed to encourage exploratory reasoning and help the model learn complex reasoning patterns during training.

Mini-o3 project address

• Project official websitehttps://mini-o3.github.io/
• GitHub repository: https://github.com/Mini-o3/Mini-o3
• HuggingFace model libraryhttps://huggingface.co/Mini-o3/models
• arXiv technical paper: https://arxiv.org/pdf/2509.07969

Application scenarios of Mini-o3

• e-commerce scenariosIt helps users quickly find their target products among a vast number of product images. On apparel e-commerce platforms, users can upload a picture to search for similar styles of clothing.
• Smart HomeIn a smart home environment, cameras capture images to help users quickly find lost items, such as keys and remote controls.
• Surveillance video analysisIt enables the rapid location and identification of specific targets in surveillance videos, such as finding specific people or items in crowded places.
• Abnormal behavior detectionIt analyzes abnormal behaviors in surveillance videos through multiple rounds of reasoning, such as intrusion detection and abnormal activity recognition.
• Navigation in complex scenesIn complex road conditions, multi-round visual reasoning helps autonomous driving systems better understand and plan routes, such as in scenarios with obstructions or complex traffic signs.