Core Viewpoint - MiroMind ODR (Open Deep Research) is introduced as a powerful open-source deep research model, achieving a GAIA test score of 82.4, surpassing other models like OpenAI's Deep Research and Manus [2][5]. Group 1: Model Performance and Features - MiroMind ODR has the highest performance score of 82.4 in GAIA validation, outperforming models such as OpenAI Deep Research (67.4) and Manus (73.3) [2][5]. - The model is fully open-source and reproducible, with all core components, data, training processes, and frameworks available for public access [4][5]. - The project team plans to maintain a monthly update schedule for open-source contributions, indicating ongoing development and improvement [5]. Group 2: Sub-Projects Overview - MiroMind ODR consists of four sub-projects: MiroFlow (Agent Framework), MiroThinker (Model), MiroVerse (Data), and MiroTrain (Training Infrastructure) [20]. - MiroFlow supports multiple mainstream tool calls and extends large language models, achieving stable reproducibility with a performance score of 82.4 on GAIA [22]. - MiroThinker is a large language model that natively supports tool-assisted reasoning, demonstrating top performance in GAIA [23]. - MiroVerse provides 147,000 open-source training datasets, focusing on community feedback and continuous updates [26]. - MiroTrain supports stable and efficient training for deep research models, covering the entire training process [27]. Group 3: Development Team and Leadership - Dai Jifeng, a prominent figure in the project, has a strong academic background and extensive experience in computer vision and deep learning, with over 80 published papers and more than 60,000 citations [32][36]. - His previous roles include positions at Microsoft Research Asia and SenseTime, and he has returned to academia as an associate professor at Tsinghua University [40][41]. - The project aims to contribute to AGI (Artificial General Intelligence) research, with a mission to create self-aware digital entities that evolve with the community [45][47].
代季峰陈天桥联手AGI首秀炸场!最强开源深度研究模型,GAIA测试82.4分超OpenAI
量子位·2025-08-09 09:53