目录Qwen3-VL-32B-InstructQwen/Qwen3-VL-8B-Instruct2. InternVideo2.5 —— 目前非常成熟InternVL2.5-26B-Instructyanziang/InternVideo3-8B-Instruct如果是在你前面这个**“视频动作识别 / 羽毛球视频分析”**场景里比较Qwen3-VL-32B-InstructQwen3-VL-32B-Instruct fp8 需要48g显存Qwen/Qwen3-VL-8B-Instruct2. InternVideo2.5 —— 目前非常成熟InternVL2.5-26B-Instruct30G显存hf download OpenGVLab/InternVL2_5-26B-AWQ --local-dir ./OpenGVLab/InternVL2_5-26B-AWQyanziang/InternVideo3-8B-Instructvideo_path your_video.mp4 fps 1 min_pixels 128 * 32 * 32 max_pixels 128 * 32 * 32 messages [ { role: user, content: [ {type: video, video: video_path, fps: fps}, {type: text, text: Please describe this video in detail.}, ], } ] processor.video_processor.size { longest_edge: max_pixels * max_frames, shortest_edge: min_pixels * min_frames, } inputs processor.apply_chat_template( messages, tokenizeTrue, add_generation_promptTrue, return_dictTrue, fpsfps, return_tensorspt, ) inputs inputs.to(model.device) output model.generate(**inputs, max_new_tokens1024, use_cacheTrue) generated_ids [o[len(i):] for i, o in zip(inputs.input_ids, output)] print(processor.batch_decode(generated_ids, skip_special_tokensTrue)[0])