Local AI video search: what runs on your own Mac in 2026
What Apple Silicon can actually do to your footage offline, how much worse it is than the cloud, and when that trade is obviously correct.
Two years ago, “AI analysis of your footage” meant uploading it. That was fine for a travel vlog and a non-starter for anything with a confidentiality clause, which is most paid work.
That has changed, quietly and quite fast. A model that describes a video clip well enough to find it again now fits on a laptop. Here is what that actually gets you, what it costs in quality, and where the cloud still deserves the job.
What “local” means here
Three separate model types run on-device now, and they are worth separating because they perform very differently.
Speech to text. The most mature by a distance. Whisper-class models run comfortably on Apple Silicon and produce transcripts at a quality that is genuinely close to cloud services, in dozens of languages, with word-level timecodes. If your work involves people talking, this alone changes your workflow. Detail in transcribing video on a Mac.
Vision and language. Models in the two to eight billion parameter range that look at frames and describe them in words. This is the newer capability and the one with the biggest gap to the cloud. It is good enough to find a clip, not good enough to write your shot list.
Classification. Small, fast, task-specific models: is this a talking clip or a cutaway, is this footage blurred, is this material sensitive. These have been fast and reliable on-device for years.
On Apple Silicon these run through Apple’s own frameworks and, increasingly, through MLX, a machine learning framework built specifically for the unified memory architecture. That architecture is the reason a laptop can do this at all: the GPU and the CPU share the same memory, so a multi-gigabyte model does not have to be copied across a bus.
What the hardware actually needs to be
Honest numbers, not marketing ones.
| Machine | What to expect |
|---|---|
| Intel Mac | Local vision models are impractical. Use the cloud. |
| M1 or M2, 8 GB | Small models only, heavy memory pressure, slow. Possible, not pleasant. |
| M1 to M4, 16 GB | The realistic entry point. Analysis around real time or faster, per clip. |
| 32 GB and up | Larger models, better descriptions, comfortable multitasking while it runs. |
| M-series Pro or Max | Meaningfully faster, since more GPU cores means more throughput. |
Also budget disk space. A vision model is several gigabytes, and a good speech model is another one to three. Downloading them is a one-time cost, but it is a real one on a laptop that is always at 90 percent full.
Throughput in practice: expect analysis of a short clip to take on the order of seconds, and a large card of footage to be a background job you start and walk away from. It is not instant. It is unattended, which matters more.
How much worse is it, really
This is the question everyone wants a straight answer to. The honest one:
Categories: about the same. Kitchen, street, beach, forest, office, crowd, interview, drone shot. A local model handles all of these.
Specifics: noticeably worse. Where a cloud model says “pufferfish over a coral reef”, a local model says “fish underwater”. Where a cloud model reads the sign and tells you the shop name, a local model says “a shopfront”.
Text in frame: much worse. Reading whiteboards, signage and screens is a strength of large models and a weakness of small ones.
Speech: nearly the same. This is the outlier. On-device transcription is close enough to cloud transcription that the difference rarely matters for clean audio.
Consistency: local is often better. A small model gives you the same vocabulary every time, which for tagging is a genuine advantage. A large model is more articulate and more variable.
Now the important part: for retrieval, coarse descriptions are usually sufficient. Your search is not “find the pufferfish”. Your search is “which of these 400 clips is the underwater one with the fish”. “Fish underwater” answers that in one second. You then use your eyes, which is what you were going to do anyway.
The cases where coarse is not enough are logging for stock libraries, archives that need consistent controlled vocabulary at depth, and any workflow where somebody downstream will search using words you did not anticipate.
When local is obviously the right choice
Not a preference, a requirement:
- Contracts with confidentiality clauses. Most client production agreements have them. An upload can breach one even if nothing bad happens to the data.
- Medical, legal, and law enforcement material. Regulated, and the regulations do not care about your workflow.
- Unreleased products, pre-embargo material, internal corporate footage. Leak risk is somebody’s actual job to worry about.
- Anything involving people who did not consent to processing. Bystanders, minors, sensitive locations.
- Personal material you simply do not want on someone else’s server.
- Volume. Ten thousand clips at any per-clip price is real money. On-device analysis costs electricity.
- No connection. Editing on a plane, in a hotel with hostile wifi, on location.
That list covers a very large share of professional work, which is why on-device capability stopped being a niche feature.
What local analysis still cannot do
To be complete about the limits:
- It does not know your people or places. No model does, unless you tell it.
- It cannot read your project. Which take is the good one is not a visual property.
- It does not index by timecode. You get a description of the clip, not a log of the clip. That distinction is covered in searching video by objects.
- It will not touch camera RAW. R3D, BRAW, ProRes RAW, ARRIRAW and camera DNG are closed containers for most tooling. Transcode first.
- It heats up your laptop. Sustained analysis of a large card is a heavy GPU job. On a fanless machine, it throttles.
A sensible hybrid
The setup most professionals converge on, once they have tried both:
- On-device by default, for the whole card, immediately after ingest. Free, private, unattended.
- Cloud selectively, for the material where precision pays: the stock-worthy B-roll, the shoot you will resell, the archive you will search in five years with words you cannot predict today.
- Transcription always local, since the quality gap is small and transcripts are the highest-value output per unit of effort.
That way the confidentiality question only arises for the clips you have consciously decided to send, and it never arises for the bulk of your library.
Where the search itself happens
Worth stating plainly: the model is not the search engine. The model writes text. Your operating system searches text.
So the local pipeline ends where every good pipeline ends: descriptions in filenames, subject tags as Finder tags, transcripts as text files next to the clips. Spotlight indexes all three. The result is that your “local AI video search” is just Spotlight, on data that finally exists.
That is a feature, not a compromise. It means the search keeps working when the model is uninstalled, the app is gone and the drive is plugged into a machine that has never heard of any of this. See finding clips without Premiere for what that looks like day to day.
Cliptag runs both modes on macOS. The on-device mode uses a vision model through MLX on Apple Silicon and is free and unlimited, with an on-device check that keeps sensitive clips from ever being sent to the cloud even when cloud analysis is switched on. Cloud analysis is there for the material where precision is worth it, with 25 analyses a month on the free plan.
