hugging-apps/timelens2-8b-temporal-grounding

TimeLens2-8B — Video Temporal Grounding

This Space runs MCG-NJU/TimeLens2-8B, a 9B video multimodal LLM (fine-tuned from Qwen3-VL-8B-Instruct) that achieves state-of-the-art on seven temporal-grounding benchmarks (average mIoU 48.0).

Usage

1. 1.Upload a video.
2. 2.Enter a natural-language query describing the event you want to locate.
3. 3.Click Ground — the model returns a JSON array of [start, end] time spans (in seconds) where the query is relevant.