omni-macos — a semantic Finder on Apple Silicon · Launch Video Breakdown: Hook, Pacing & Motion Design

Scene-by-scene timeline & spoken transcript

1. 0:00 – 0:07

   The Hook

   History of Semantic Search

   “Semantic search has existed for decades.”

   On screen

   Semantic search has existed for decades. 1990 LSI 2003 LDA 2013 word2vec 2018 BERT 2020 vector DBs

   Camera

   Static top-down planar view

   Motion

   Sequential tag dropping with string physics

2. 0:07 – 0:15

   Problem Agitation

   Cloud Privacy Vulnerability

   “Today it is mostly a cloud service, so your files are uploaded and searched on remote servers.”

   On screen

   Today it is mostly a cloud service. Your files are uploaded to remote servers. Uploading 259 files Embeddings API Vector DB Search API RAG platform

   Camera

   Fixed frame

   Motion

   Paper cloud cutouts with suspended tags and floating photos

3. 0:15 – 0:25

   Product Reveal

   100% Local On Mac

   “Omni keeps all of it on the Mac that holds the files, with no server and no network connection.”

   On screen

   Everything stays on your Mac. It needs no server and no network.

   Camera

   Vertical downward pan through layered strata

   Motion

   Multi-plane layered cardboard parallax

4. 0:25 – 0:29

   Product Reveal

   Brand & Mascot Reveal

   “This is Omni.”

   On screen

   Omni Semantic search that runs on your Mac.

   Camera

   Static centered lockup

   Motion

   Paper cutout burrow reveal with mascot pop-up

5. 0:29 – 0:37

   Feature Teaser

   Swift and Metal Engine

   “Running the embedding model locally is straightforward. Omni runs it in Swift on Metal, without Python.”

   On screen

   Running the model locally is straightforward. Swift Metal No Python 83,105 tokens/s indexing throughput on an M3 Ultra

   Camera

   Fixed graphic focus

   Motion

   Animated metric counter with matrix tile fill

6. 0:37 – 0:52

   Problem Agitation

   Index Maintenance Complexity

   “Inference, however, is only one part of the problem. The harder part is keeping the index current when folders hold version 1, version 2, final, and final final.”

   On screen

   Inference is only one part of the problem. The harder part is keeping the index current. Folders hold many versions of one document. report_v1.pages report_v2.pages report_final.pages report_final final.pages

   Camera

   Downward shift into subterranean workspace

   Motion

   Paper document fanning and duplicate layer stack

7. 0:52 – 1:03

   Feature Teaser

   Content-Addressed Deduplication

   “Omni stores each passage once, keyed by its content. A copy adds no vectors, and an edit re-encodes only the passage it changes.”

   On screen

   stored once Four files with 24 passages need 8 vectors. A copy adds no vectors. Saving an edit takes 8.5 ms. median on an M3 Ultra

   Camera

   Static diagram view

   Motion

   Color-coded chunk highlights linking to vector slot bins

8. 1:03 – 1:13

   Feature Teaser

   Background Indexing Throttling

   “Indexing shares the machine with you. While you type, it gives the GPU smaller units of work, within one memory limit that you set.”

   On screen

   Indexing shares the machine with you. While you type, the indexer works in smaller units. The longest wait for a query drops from 3.1 s to 1.4 s. indexing your memory limit

   Camera

   Horizontal profile view

   Motion

   Split-chunk animated bar progression and live limiter needle

9. 1:13 – 1:24

   Feature Teaser

   1-Bit Quantization Scan

   “There is no vector database. Each query scans a 1-bit copy of every vector and rescores the best candidates exactly.”

   On screen

   There is no vector database. exact vector 1-bit copy, 16x smaller. The best candidates are rescored exactly. It recovers 95.8% of the exact top 10. #1 #2 #3 #4 #5 #6 #7 #8 #9 #10

   Camera

   Static analytical layout

   Motion

   Binary scan wave filtering into top-10 exact rescored cards

10. 1:24 – 1:32

    Feature Teaser

    Local MCP Agent Support

    “Local agents can search the same index over MCP, and their queries never leave the machine.”

    On screen

    Local agents search the same index. agent MCP search_files("q3 board deck") 12 results · 9 ms agent MCP find_similar("logo_final.png") 40 similar · 6 ms agent MCP search_files("Rechnung März") 7 results · 10 ms served on localhost only

    Camera

    Centered mascot anchor

    Motion

    Terminal tooltip popups linked by conduit paths

11. 1:32 – 1:43

    Feature Teaser

    Academic Validation

    “We describe the design and its measurements in a paper at the NeurIPS 2026 Workshop on On-Device Intelligence.”

    On screen

    On-Device Omni-Modal Search on Apple Silicon The design is published. NeurIPS 2026 Workshop on On-Device Intelligence 9.7 ms a text query 8.5 ms saving one edit 83,105 tokens/s indexing 5 Macs M2 to M3 Ultra

    Camera

    Overhead tilt settling onto paper sheet

    Motion

    Paper research doc drop with embossed medal and stats

12. 1:43 – 2:11

    Product Reveal

    Native macOS Interface Walkthrough

    “In the app, you search by meaning in any language, find similar files, and browse folders as you do in Finder.”

    On screen

    Search 427,560 files Search by meaning. mountain lake at sunrise type:image Vollmond über dem Meer Search in any language. Find similar files. similar to: figure_19.png Browse folders as in Finder. Replay past searches. cat sleeping on a sofa

    Camera

    Push-in to cardboard laptop screen, followed by flat screen UI record

    Motion

    UI screen recording with live typing, filtering, and tag chips

13. 2:11 – 2:25

    Call to Action

    Open Source Distribution

    “Omni is free and open source.”

    On screen

    Omni Free and open source. hanxiao.io/omni macOS 14+ Apple silicon Apache 2.0

    Camera

    Centered final lockup to subtle fade-out

    Motion

    Paper mascot wave with smooth badge pop-ins