小米MiMo-VL VS 千问Qwen2.5-VL | 多模态模型实测

Core Viewpoint - The article discusses the performance of Xiaomi's MiMo-VL-7B multi-modal model, highlighting its strengths and weaknesses compared to the Qwen2.5-VL model, particularly in various testing scenarios. Group 1 - MiMo-VL-7B model outperforms several multi-modal understanding models, especially Qwen2.5-VL, in various tests [3][5]. - The testing results indicate that the SFT (Supervised Fine-Tuning) and RL (Reinforcement Learning) versions of MiMo-VL-7B show similar performance, while the "think" version significantly outperforms the "no-think" version [5][6]. - MiMo-VL-7B's performance in recognizing handwritten OCR is noted to be poor [5][9]. Group 2 - In table recognition tasks, MiMo-VL-7B's "think" model performs well, while the "no-think" model and Qwen2.5-VL struggle [9][10]. - For medium complexity tables, MiMo-VL-7B-SFT "think" model approaches correctness, while other models fail [18][19]. - The article emphasizes that MiMo-VL-7B-SFT "think" model shows better results in complex table recognition compared to its counterparts [26][27]. Group 3 - The article concludes that Xiaomi's MiMo-VL model is impressive overall, particularly the "think" model, which excels in most capabilities except for handwritten OCR [67][68]. - Despite its strengths, the article suggests that the claims of MiMo-VL-7B significantly outperforming the 72B model may be exaggerated [68].