Find any moment in a video by describing it

Upload a video and type what you're looking for: Narendra Modi, a red car, people dancing. EmbeddingGemma 2 scans the frames and marks every segment where it appears.

Choose a video or drop one anywhere on the page

Private. Everything runs on your GPU. The video is never uploaded.
Light. ~300 MB download (4-bit), cached after the first visit. The audio encoder is never loaded.
Fast to re-search. Frames are indexed once; every query after that is instant.

similarity match not indexed yet

Preparing…

Change video

Indexing settings

Sample a frame every Frame detail Vector size (MRL) Model precision

Precision: Google's guidance is to run this model in bfloat16 or float32 and not float16, whose range its activations exceed. The ONNX fp16 build used here measures 0.9998 cosine similarity to float32 (and matched 4-bit within 0.002 in testing), so it is used for the CPU path to halve the download — pick fp32 if you want the safest option.

Memory

Model weights: not loaded
Frame index: –
Frames in flight: 0
Thumbnails: 0
JS heap: n/a

The video is streamed from disk, never copied into memory. Frames are resized on the GPU, embedded, and freed right away. Only small vectors are kept.