Underdog shows off a 32MB speech model in new demo

AI startup Underdog has released a browser game called Woof to showcase its new frontier text-to-speech model, which weighs in at roughly 32MB — small enough, the company says, to run on most devices without an internet connection. In the demo, players type text on an on-screen typewriter and hear it spoken back, with all voices generated by Woof.

The game's setting reinforces the pitch: a retro-tech room with a typewriter, CRT monitor and desk lamp whose switch produces a click sound. Its intro warns players about "intrusive" cloud providers in town, an apparent jab at cloud-hosted models. A tuner control lets players generate different kinds of music, underscoring Underdog's claim that the model produces both speech and music. The model has 31 million parameters and is open for anyone to download.

Underdog has framed Woof as one piece of a broader edge-computing strategy. Weeks earlier, the company launched Bark, a 167MB speech-to-text model aimed at transcription and dictation on edge devices. It has also released a 27-billion-parameter model that can run on laptops, plus a waitlisted app positioned as a locally-hosted alternative to Muse and Instinct.

Founder Sigil Wen told TechCrunch that he began building the system after deciding against notetakers like Granola, preferring AI models that run on users' devices so people can own their data. "I think we will be able to run most capable models of today on the devices you already own within months," he said, adding that the company wants to train open, American-made models that enable private AI, citing his manifesto.

Underdog positions Woof toward local assistant experiences and personalized podcasts, though it acknowledges a 32MB model likely won't replace larger systems. What remains to be seen: whether a 31-million-parameter model delivers voice quality good enough for daily assistant use, whether browser-based inference holds up beyond a demo, and whether Wen's timeline for running today's most capable models on consumer devices proves accurate.