Minicpm-V-4

https://preview.redd.it/bpy1hvrqoihf1.png?width=3835&format=png&auto=webp&s=fe9dbeabaa53cb1fcefaeb927e53d9042ee39e8a Among the three 4B-level VLMs(using Q6 GGUF for Minicpm, F16 weight & vllm for Qwen, Q4 Ollama GGUF for gemma), I still think that Qwen2.5-VL-3B performs relatively better in extracting structured information from images. However, I'm particularly interested in this model's video understanding capability. Given its high token density—encoding a 448×448 image into a single tile of 64 tokens, meaning each token represents approximately 3,000 pixels—it could be a promising candidate for training a compact video understanding model.