latent arranges your images, text, code, and PDFs in 3D based on what's in them. Search the scene, inspect clusters, and switch projection models without sending your files anywhere.
macOS 26+ · Apple Silicon · Free tier available
You give it files. It embeds their content locally, then gives you a spatial map you can search, filter, and inspect.
TinyCLIP is bundled for fast local embeddings. SigLIP is available as an optional higher-quality model download. Your files never touch a server.
Drop images, text files, source code, PDFs, or whole folders into the same space. latent imports supported files and places them relative to each other.
Type a natural-language query like "rose" and matching files stay emphasized while the rest fade back.
Use PCA for a fast overview or t-SNE for tighter clusters. Adjust t-SNE parameters and scene spread directly from the app.
Browse imported files in the side navigator, filter by name, select items in the scene, and inspect previews and metadata in the inspector.
The 3D scene is rendered with Metal, so orbiting, selection, search dimming, and dense thumbnail clouds stay responsive.
Drag anything onto the window — images, text, code, PDFs. Mix types freely.
TinyCLIP processes each file on-device and generates an embedding. You can optionally download SigLIP for higher-quality embeddings.
Vectors are reduced to 3D and rendered in real time. Search the scene, tune the projection, click items to inspect them, or select groups directly.
Version 1.2 adds the current app workflow: file navigation, semantic scene search, model controls, and projection tuning.
Search by meaning, not just filename. Matches remain visible in the 3D scene while unrelated items fade.
Switch between PCA and t-SNE, then tune perplexity, iterations, and learning rate for the shape of the map.
Free tier has no time limit. Pro is a one-time purchase. No subscription, because there's nothing running on our end.
An embedding is a list of numbers that represents what a file "means." A photo of a dog and the text "golden retriever" would have similar numbers, so they'd appear near each other in the visualization. TinyCLIP is bundled, and SigLIP can be downloaded as an optional higher-quality model.
Yes. The Mac app does not track usage and does not upload files, embeddings, or searches. The website uses OpenPanel for basic page analytics, but the app itself runs locally.
The free tier supports 300 items. Pro removes that limit, though performance depends on your hardware. We've tested with 1,500+ items without issues.
Images (JPEG, PNG, WebP), plain text, source code, and PDFs. You can mix them all in the same space, including files imported from folders.
Everything runs on your Mac, so there's no ongoing cost on our side. If we add cloud features later, that'd be separate.
macOS 26 or later, on Apple Silicon. Intel Macs aren't supported. The neural engine is what makes on-device inference fast enough to actually use.