Mini-o3 – a visual reasoning model launched jointly 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 tackle complex visual search problems. The model uses reinforcement learning and image-based tools to perform deep, multi-round reasoning, with reasoning steps scalable to dozens of iterations. Mini-o3 enhances reasoning ability and interaction depth through the creation of challenging datasets, iterative data collection processes, and an over-turn masking strategy. It achieves state-of-the-art performance on multiple visual search benchmarks. All code, models, and datasets are open-source, enabling reproducibility and further research.

Key Features

• Multi-round interactive reasoning: Performs deep, multi-round reasoning, scalable to dozens of iterations, solving complex visual search problems through step-by-step exploration and trial-and-error.

• Diverse reasoning modes: Supports various reasoning strategies, including depth-first search, trial-and-error, and target maintenance.

• Challenging visual search: Accurately locates and identifies targets in high-resolution images even when the targets are small and surrounded by numerous distractors.

• Excellent performance: Achieves state-of-the-art results on multiple visual search benchmarks (e.g., VisualProbe, V* Bench, HR-Bench, MME-Realworld), demonstrating strong visual reasoning capability.

• Open-source: All code, models, and datasets are open-source, allowing researchers to reproduce results and advance related research.

Technical Principles of Mini-o3

• Cold-start Supervised Fine-tuning (SFT): Uses a small number of hand-crafted examples to generate high-quality, diverse multi-round interaction trajectories with a visual-language model (VLM) possessing strong contextual learning ability.

• Reinforcement Learning (RL): Uses an over-turn masking strategy to avoid penalties when interaction rounds exceed limits, allowing natural extension to dozens of rounds during testing.

• Lower Down Max Pixels: Reduces the maximum pixel count per image, increasing the number of rounds allowed per interaction and improving the model’s ability to solve long-horizon problems.

• Challenging dataset (Visual Probe Dataset): Constructs a dataset of thousands of visual search problems designed to encourage exploratory reasoning, helping the model learn complex reasoning patterns during training.

Project Links

• Official site: https://mini-o3.github.io/

• GitHub repository: https://github.com/Mini-o3/Mini-o3

• HuggingFace model hub: https://huggingface.co/Mini-o3/models

• arXiv paper: https://arxiv.org/pdf/2509.07969

Application Scenarios for Mini-o3

• E-commerce: Helps users quickly locate target products among massive image collections. For example, on a fashion e-commerce platform, users can upload an image to find similar clothing items.

• Smart home: In smart home environments, captures images via cameras to help users quickly find lost items such as keys or remote controls.

• Surveillance video analysis: Rapidly locates and identifies specific targets in surveillance footage, such as finding a person or object in crowded areas.

• Anomaly detection: Uses multi-round reasoning to detect abnormal behavior in surveillance videos, such as intrusions or unusual activities.

• Complex scene navigation: Assists autonomous driving systems with multi-round visual reasoning to better understand and plan routes in complex road conditions, such as scenes with occlusions or complicated traffic signs.