AI
Underdog debuts tiny 32MB voice model that runs offline on your device
The AI startup is pushing back against cloud-dependent models by releasing compact, edge-ready speech technology designed to work without internet.
TechCrunch — AI
While the AI industry focuses on scaling massive models that demand enormous computing resources, a contrarian wave of startups is taking the opposite approach—building stripped-down systems that can operate directly on personal devices. Underdog, a new player in this space, is making that philosophy tangible with its latest offering.
The company has rolled out an interactive browser game showcasing Woof, its new voice generation model that weighs just 32MB. The playful demo features a retro-inspired office setting with a typewriter, old-school tuner, and CRT monitor—complete with a tongue-in-cheek warning about "intrusive" cloud providers. Users can type text into the typewriter and hear it spoken back, or adjust the tuner to generate different musical passages. Everything runs locally in the browser without requiring an internet connection.
A compact model with real utility#
Woof contains 31 million parameters and is freely available to download. While small models like this one cannot yet match the sophistication of their larger counterparts, they open up new possibilities: powering private voice assistants, generating personalized audio content, or enabling transcription directly on users' hardware.
The timing follows Underdog's earlier launch of Bark, a 167MB speech-to-text model also designed for offline, edge-device operation. The startup has additionally developed a 27-billion-parameter model capable of running on laptops and is preparing a waitlisted application that would serve as a locally-hosted rival to existing cloud-based services.
The vision behind the company#
Founder Sigil Wen framed the initiative as a reaction to existing note-taking tools and AI services that offload processing to external servers. He wants users to retain ownership of their data and processing by running capable models on hardware they already own. In a statement to TechCrunch, Wen argued that "we will be able to run most capable models of today on the devices you already own within months," emphasizing his push for openly available, American-developed AI that prioritizes privacy. Underdog has published a manifesto outlining these principles, positioning itself as an alternative to cloud-first competitors.

