Best match
Wide two-shot, café. She laughs; he looks away.
Why it matched
- Wide shot
- Laugh follows the line
- No crew in frame
Multi-modal intelligence
for real-world video
The whole picture. The exact moment.
We're building pr0metheus, a video-native AI model with one target: beat Gemini at understanding real-world video.
The problem
There is no grep for video.
Text is searchable by default. Video is not. Today's tools are fast and shallow, or smart and slow.
$ grep -n "dog" script.txt 14:The dog runs past the café. $ grep -n "dog" A001_C004.mov $
The dog is right there at 00:02:41. To grep, it's just pixels.
- Fast, shallow
- One vector per clip. “Man with dog” ≈ “dog without man.”
- Smart, slow
- Large models read every clip. Too slow for a whole library.
Model 01
In developmentpr0metheus
Built for footage, not frames.
pr0metheus is a vision-language model built on open weights, designed to follow who does what, in what order, to the second. Real footage is hours long, messy and multi-camera. That's where we're building it to beat Gemini.
Built to win at
- Negation “no crew in frame”
- Event order “laughs after the line”
- Exact in/out 00:14:32–00:14:39
- Whole library every drive, every project
- Private on-device first
Editorial intelligence
Describe the moment. Get the exact seconds.
Plain-language search across a whole library. Time ranges, not files, each with the reason it matched.
Illustrative footage and results. pr0metheus is in development.
Concept preview with illustrative results. The search “wide shot, she laughs right after ‘I told you so’, no crew in frame” returns 00:14:32–00:14:39, 7 seconds, in clip A014_C007: wide two-shot, café. She laughs; he looks away. Why it matched: wide shot, laugh follows the line, no crew in frame. The search “the clean take where he stands up, then drops the mug” returns 00:27:04–00:27:09, 5 seconds, in clip B003_C012, take 4: medium shot, kitchen. He stands; the mug slips and shatters. Why it matched: stands, then drops; shatter on audio; clean delivery. The search “empty street at dusk, ‘OPEN 24H’ neon, from last year’s shoot” returns 00:03:18–00:03:26, 8 seconds, in clip A002_C021, archive 2025: slow push-in down an empty street. Neon glows at dusk. Why it matched: sign reads OPEN 24H, no people, shot in 2025.
How it works
From raw video to ready-to-publish.
- 01
Detect
Split footage into shots, faces, words, sounds and on-screen text.
- 02
Understand pr0metheus
Read picture, speech and sound as one, across time.
- 03
Select
Rank every candidate. Keep the best take, with exact in and out.
- 04
Edit
Send selects to the timeline, ready to cut.
Under the hood
Fast where it can be. Smart where it counts.
Cheap retrieval casts a wide net. pr0metheus reranks and verifies the few that matter.
- Retrievewhole library
- Rerankpr0metheus
- Verifynegation, order, in/out
Why now
Accuracy just got cheap.
Published research
54.5 → 87.1
Text-to-video Recall@1 on MSR-VTT, zero-shot: prior best model vs. training-free VLM reranking.
Vote-in-Context, 2025<1%
of full reranking runtime once visual work is cached and reused, keeping >96% of performance.
miniReranker, 20262025–26
Premiere, Resolve and Final Cut shipped basic AI search. Hard queries remain open.
Vendor release notesRoadmap
From finding to finishing.
The model
Train pr0metheus and measure it against Gemini on an editor-grade benchmark.
The search
Early access for editors: exact-moment search across whole libraries, on-device.
The cut
Selects become first cuts. Raw video in, ready-to-publish out.