The field of long video understanding is moving towards more efficient and scalable solutions, with a focus on reducing the overwhelming volume of visual tokens generated from extended video sequences. Recent developments have introduced innovative methods for keyframe selection, visual token compression, and dualistic visual tokenization, which have shown significant improvements in accuracy and efficiency. These advancements have the potential to enable multimodal large language models (MLLMs) to process long videos in a more practical and effective manner. Noteworthy papers in this area include FOCUS, which proposes a training-free and model-agnostic keyframe selection module, and FLoC, which introduces an efficient visual token compression framework based on the facility location function. Additionally, the Wave-Particle dualistic visual tokenization approach has shown promising results in unifying understanding and generation within a single MLLM. Other notable works include the introduction of the trigger moment for grounded video QA, the development of a unified benchmark for visual token pruning, and the release of new models such as NVIDIA Nemotron Nano V2 VL. These developments demonstrate the rapid progress being made in the field of long video understanding and highlight the potential for future innovations.