Most AI announcements this year have been about scale. Underdog went the other direction and made the size the demo.
The startup released a browser game called Escape built on Woof, a text-to-speech model that occupies 32MB and holds 31M parameters. Everything inside runs locally, with no server call and no connection required on most devices.
The scene is a retro room with a typewriter, a CRT monitor and a desk lamp that clicks. Type a sentence and the game speaks it. Rotate a tuner and music shifts. An opening warning jokes about intrusive cloud providers.
Small models, owned data
Founder Sigil Wen said the project started from a personal annoyance. He wanted a note-taking system that did not ship his audio to a vendor, so he built one from models that live on the machine.
His claim in an interview with TechCrunch was blunt: within months, most of today’s capable models will run on hardware people already own. Underdog also published a 167MB speech-to-text model called Bark weeks earlier, aimed at dictation, plus a 27-billion-parameter model it says runs on laptops.
The company pitches itself as a local alternative to hosted assistants from Meta and OpenAI.
The catch is capability
Nobody would mistake a 31M-parameter generator for a frontier speech system. Prosody flattens, unusual names get mangled, and emotional range is thin. Woof will not narrate an audiobook.
That is not the bet. The useful territory for a model this small is mundane: reading a notification aloud, voicing a form field, generating a short podcast from a document, giving a local assistant something to say without a network hop.
The economics are hard to argue with. Inference on a phone costs a fraction of a cent in battery and nothing in fees, and no data crosses a border.
