What this is
In late September, Google DeepMind open-sourced EmbeddingGemma 2—an embedding model that converts text, images, audio, and video into comparable numeric sequences, weighing in at just 865MB. This week, indie developer ARahim3 released DigUp on GitHub, a native macOS app that packs this model into system-level file search.
Usage is straightforward: describe what you remember in natural language, and the app jumps straight to that frame or passage. Phrases like "the moment a zebra appears in the video," "the clause about pets in the contract," or "that photo of the dog on the beach" all land hits—even searches in Bengali or Arabic work. Code files get function-level targeting.
Under the hood: GGUF quantization + llama.cpp + Apple Metal acceleration, written entirely in Swift with zero Python dependencies. Search loads only a 250MB text encoder; results arrive within 0.1 seconds of stopping input. Indexing peaks under 2GB, and the helper process exits automatically once the index is built. All data stays local; the app only touches the network to download the model, and update checks can be disabled.
Hard limits: Apple Silicon only (M1 and up) + macOS 14+. MIT license, free.
Industry view
What we think deserves credit: Google has long pursued the "small model" path—the Gemma line was designed to run on phones and laptops. The fact that a developer assembled a full app in days signals the gap between model layer and application layer is shrinking. An 865MB small model + a minimal wrapper = a real tool, a combo we hadn't seen in the past two years.
What we think deserves a cool head: First, Apple Silicon-only support is a hard gate—the majority of Mac users globally and almost all Windows users are locked out. Second, the demo scenarios ("zebra frame," "pet clause") are carefully curated examples; hit rates on messy folders and fuzzy memories remain unverified. Third, this is a Reddit-based personal project with no company backing—updates and long-term support are open questions. Local AI's battery drain and first-time indexing time cost also go unmentioned in the demo.
Impact on regular people
For enterprise IT: Files never leaving the device carries real weight for compliance-heavy industries like finance, healthcare, and legal. Local AI gives "AI assistants on the job" its first viable compliance path.
For working professionals: Going forward, your accumulated notes, recordings, and contract PDFs won't have to rely on file names—describe the content and find it. Caveat: you need an Apple Silicon Mac.
For the consumer market: AI is no longer just ChatGPT in a browser or the voice assistant on your phone. We're starting to see AI "living inside the OS"—a trend that will push PC vendors to re-emphasize AI compute specs.