INDEPENDENT EVIDENCE / UPDATED OCTOBER 8, 2026

EmbeddingGemma 2 Browser Video Search Benchmarks

Useful matches. Real misses. Measured waiting times.
This is a small lab dataset, not a general accuracy claim.

By FindVideoMoment · Updated October 8, 2026. This EmbeddingGemma 2 video search benchmark reports independent POC measurements and their method, separately from voluntary community data.

Independent laboratory results

Measured October 7–8, 2026 on Apple M5, 16 GiB RAM, macOS 26.5.1, Tabbit Chromium 152, Apple Metal WebGPU. Transformers.js 4.3.1, q4, model revision daa72c51243991dfcaf9f9137d2c573d8f7790c0, 1fps and four-second windows. These are prior POC results, not release-build measurements.

Video durations: LibreOffice 14:19; Blender 22:50. The 70-token runs were 39.9% and 51.1% faster than their respective 140-token baselines. The quality protocol checked whether a candidate window’s center fell inside one pre-labelled target interval. Labels were not independently reviewed and did not enumerate all valid alternative scenes.

At 70 tokens, LibreOffice indexing took 491.90 seconds. Top-3 retrieval found 7/15 targets: 7 of the 15 pre-labelled LibreOffice targets had a matching candidate among the first three results. This is a result for that frozen query set, not general model accuracy.

Failures matter

LibreOffice’s landscape print preview, cell-border popup and selected four-column table were missed in the Top-5 at 70 tokens. Blender’s faceted cylinder, selected lower semicircle and upright green leaf were also missed. The dataset does not establish dependable professional retrieval.

Synthetic stress tests — 140 tokens

No crash or sustained linear JS-heap growth was observed in these runs. GPU memory and total-process peak memory were not verified. A 300-second run previously encountered Apple hardware decode error NSOSStatusErrorDomain -12905; its retry succeeded. A 70-token, 60-minute release-build stress test remains a launch gate.

First use, progressive search and reuse

• POC-1 cold model assets: 315.82 MB and 130.75 seconds for loading plus initialization. Cached reload: 2.43 seconds. One environment, not a universal estimate.
• POC-1 160-second animated action clip: 121.00 seconds indexing, Top-1 3/10 and Top-3 8/10. A repeated 300-second clip took 260.68 seconds in video mode.
• POC-2 first searchable window: 1.86 seconds. One preselected, previously known early query produced a useful Top-3 at 121.82 seconds. This is not a general time-to-first-match estimate.
• The saved 60-minute 140-token index contained about 7.263 MB of vectors and thumbnails. One reselect-and-restore measurement took 1.345 seconds, excluding model load, with zero repeated visual inference.

Privacy scope of the original experiments

The observed browser network traffic did not upload video, frames, thumbnails, queries or embeddings. The wider research process was not strictly zero-upload: thumbnail contact sheets were sent into an assistant session for annotation. Browser-runtime privacy and the entire research workflow are different scopes.

Voluntary community statistics

Rolling window: last 30 days. Snapshot checked: 2026-10-08T03:54:59.780Z. Refresh target: hourly.

How community data is counted

Only consented sessions that actually start indexing qualify. We count the first started run per temporary tab session and its first terminal event. Completion rate uses started sessions as its denominator; median indexing ratio includes only completed runs, creating survivor bias. Duplicate event IDs are ignored, fields and impossible timing ratios are rejected, and global plus session limits restrict spam. These are self-selected, unverified browser reports, not unique people or laboratory ground truth.

Groups below 30 qualifying sessions are hidden; completion metrics also require at least 30 completed sessions. User-reported found rate is Yes/(Yes + No) among explicit respondents in the consented session subset. Unsure answers are excluded from that denominator. It is not Top-3 accuracy. Answers submitted without diagnostic consent are not linked to session metrics. No comments or individual records are public. Collection remains off until retention cleanup is verified. See retention and privacy.

Google’s Video Moments Finder vs. this browser benchmark

Google’s Video Moments Finder runs inside the Google AI Edge Gallery app on Android and iOS, as described in the official Google announcement. FindVideoMoment is an independent browser tool: it searches local video without uploading it and can run on compatible desktop Chromium browsers with WebGPU. It is not developed, operated, certified or endorsed by Google. Both offer on-device search; the distinction here is the app versus browser experience, not a claim of better accuracy or speed.

These timings and retrieval counts measure our independent browser POC, not Google AI Edge Gallery. No equivalent test of the official app was run, so this dataset cannot rank the two tools against each other.

Sources and limitations

Prior POC reports and frozen evaluation files are preserved in the independent local research repositories. No copyrighted video frames are republished on this page. The animated action sample used Big Buck Bunny, © 2008 Blender Foundation, CC BY 3.0. Tutorial sources: Creating Charts With LibreOffice by DistroTube and Blender Tutorial. A cup of tea. by Fun with Open Source, both CC BY 3.0 according to the preserved source/license records checked October 7, 2026. No tutorial media is hosted here.

One Mac, two tutorials, small frozen query sets, no independent human adjudication, and limited codec coverage. No comparable competitor evaluation has been performed. Search demand and future Google or ChatGPT visibility remain unknown.