VideoLLM runs live video QA at 2 FPS

Most video‑large language models still operate on pre‑recorded clips, pausing after each inference. The emerging expectation that a model can watch a live feed and answer questions instantly has remained out of reach—until a system demonstrated continuous processing on a streaming pipeline. Earlier streaming attempts treated the visual front‑end and the language back‑end as separate stages, often limiting interaction to caption‑style narration or relying on explicit triggers before a response. T