Background · Part 1 of the series
Most AI tools send your data to a data centre and return an answer. Colibri does the opposite: the AI is installed on your computer and the data stays where it is. This article explains what that means in practice, why it makes the difference in media production in particular, and where the limits are.
On installation, Colibri downloads more than 25 small AI models to your computer. Each specialises in one particular task: describing images or video clips, recognising faces and text, transcribing speech and telling speakers apart, classifying sounds, transcribing and translating interview sound bites, summarising content or finding similar motifs. Together they analyse every photo, video, recording and document in your collection - on your own processor and your own graphics card.
After this initial download, Colibri works without an internet connection. Analysis, search and asking your collection run entirely on the device. No server on our side ever sees a file. When several people work together, a connection server only helps the devices find each other. The content travels directly from device to device, end-to-end encrypted.
The volumes. A shooting day easily produces several hundred gigabytes; a documentary shot over months produces many terabytes of rushes, interviews, sound and versions. Uploading four terabytes over a 100 megabit line takes almost four days before a cloud AI has seen a single file. With Colibri, the analysis starts locally - where the material already is: on the editing workstation, the NAS or the production's shared storage.
Unreleased material. Until broadcast or premiere, footage is confidential. Many commissioning clients specify in production contracts where material may be stored and who may access it. Uploading to an AI service is then not a technical question but a contractual one, and the answer is often no. With Colibri it never comes up.
People in front of the camera. Interviewees, protagonists, passers-by: whoever was filmed agreed to appear in the film, but that does not automatically mean they agree to have their face and voice processed by a third-party provider. That is why face recognition and transcription in Colibri run on your hardware, and the results stay with you.
Working where you shoot. On set, in the OB van or on the road there is often no fast network. Colibri still searches and indexes the material - right on the laptop used for editing and logging.
Originals stay originals. RAW photos, log footage from the major camera brands and large video files stay unchanged in your folders. Colibri only adds previews, descriptions and a search index alongside. With the panels for DaVinci Resolve Studio and Premiere Pro, which are nearly ready, you bring hits straight into your editor.
Archives hold material whose rights holders never agreed to a cloud upload. Publishers own unpublished manuscripts, licensed illustrations and contracts. At the Frankfurt Book Fair, the same points came up again and again: this content should not leave the building, the AI should run on your own machines, and instead of yet another silo, what is needed is a knowledge layer on top of the existing systems. That is exactly what Colibri is built for.
Because the material never leaves your systems, there is no transfer to third countries, no processing on someone else's server and no provider reading along. For your privacy notice and your record of processing activities, this means: the processing happens with you, on your hardware, under your rules.
Colibri does not train any model on your files. What it creates are descriptions, transcripts and search indexes in a Colibri database on your computer. Transcripts, subtitles and metadata can be exported in open formats. Delete them and nothing remains. The GDPR calls it the right to erasure; here it is one click instead of a request to a provider.
The same applies to copyright. Licensed archive footage, agency photos, music and style guides stay within the licence you already have, because they are not copied anywhere.
Local does not mean that everyone works on their own. Teams share a collection with roles and invitations. Tags and albums sync between devices within seconds. If the material sits on a network share, a server, in SharePoint or OneDrive, Colibri reads it where it is and copies nothing. With SharePoint, each person sees in Colibri only what they may see at the source.
For organisations that want to use Colibri in the browser, access through their own hub is in preparation: Colibri then runs on a computer or server in-house and the team reaches it in the browser. Here, too, the processing stays with you.
Large cloud models answer every request with hundreds of billions of parameters, in data centres with enormous power and cooling needs. Small, specialised models need a fraction of that. A laptop can run them, and it does so only once: every file is analysed a single time, and every later search and question costs almost nothing.
That is why Colibri runs on hardware that productions already have. A current Mac with Apple silicon or a Windows PC with a graphics card is enough for daily work. A collection of 59,000 files on a network share opens in about eight seconds.
Cloud AI is billed per minute, per file, per request or per month, and the bill grows with the archive. Colibri is a licence you buy once. There are no per-request costs, no per-terabyte costs and no subscription that cuts off access to your archive when it lapses.
Your independence goes beyond price. Colibri does not depend on a single AI provider's interface, so a change in someone else's terms does not change your workflow. Your files stay in your folders and formats, and your existing systems keep running. Colibri is a layer on top, not a replacement, and the layer can be removed at any time.
The first analysis takes computing time. A few seconds per image and document, more for video and audio, and a large collection takes correspondingly long on the first pass. After that, every search is instant. If you want to index a lot of material at once, use a powerful machine for it, such as a Mac with an M Max chip or a PC with an NVIDIA RTX card, and let the other devices benefit from its results.
And one more thing: local models know only your material. For Colibri that is intended: answers come solely from your files and cite their evidence. What is not in them, Colibri does not say. That is also the subject of one of the next parts of this series.
Only once, to download the AI models after installation, and later for updates. Analysis, search and asking your collection run fully offline. Working as a team and using sources such as SharePoint naturally need a connection to those devices and services.
No. Colibri does not train any model on your files. It only creates metadata and search indexes on your computer, and both can be deleted without a trace at any time.
Colibri reads it where it is and copies nothing. The analysis runs on your computer; the collection holds only links and the results. With SharePoint, each person's own permissions apply.
For searching and browsing an analysed collection, a Mac with Apple silicon (macOS 14 or later) and 16 GB of memory or a Windows 10/11 PC with four cores and 16 GB of RAM is enough. Importing and analysing large collections is faster with 32 GB or more and a dedicated graphics card; for production speed we recommend a Mac with an M Max chip or a Windows PC with an NVIDIA RTX card.
Colibri uses more than 25 small models, each specialised in one task such as scene description, speech recognition or translation. For these tasks they reach the quality of large general-purpose models, at a fraction of the energy and without a single file leaving the building.
Yes. Colibri reads file servers, network shares and cloud folders where they are and adds a knowledge layer on top. Nothing is migrated. Metadata, transcripts and subtitles can be exported in open formats, and panels for DaVinci Resolve Studio and Premiere Pro will soon bring hits straight into your editor.
Coming next: How does Colibri work? From import through analysis to the answer, and what runs where. After that: data protection in media work, and evidence instead of confidence.
See It Live
The beta is running. Bring a question about your archive, and Colibri answers it on your hardware.