A 32MB voice model that runs in the browser

Underdog's Woof demo turns a retro game into a test of what fits on a local device.

Underdog has released Woof, a text-to-speech model small enough to run in a browser without an internet connection. The company built a retro-themed game around it: type on a typewriter and the model speaks the text back; turn a dial and it generates music. The whole package is 31.7MB.

Woof has 31 million parameters and is open for download. It will not match larger speech systems, but Underdog argues that small models can power private assistants, offline tools, and personalized audio. The demo follows Bark, a 167MB speech-to-text model released weeks earlier.

Founder Sigil Wen says the goal is to let people run capable models on hardware they already own, keeping data local. That is a narrower promise than replacing frontier systems, but it fits devices where latency, privacy, or connectivity matter.