Underdog published a browser‑playable demo game that showcases its new, compact speech generation model called Woof. According to the company, Woof is only 31.7 MB in size (commonly rounded to 32 MB in coverage), contains about 31 million parameters, and can run locally on most devices without an internet connection.
The game and how Woof is used
The demo places the player in a retro tech–themed room with a virtual typewriter, a CRT‑style monitor, and a desk lamp that makes a sound when toggled. The intro warns about “intrusive” cloud providers in town, an apparent dig at cloud‑based models. Players can type anything on the virtual typewriter using their keyboard and the game will speak the entered text aloud via Woof. A music tuner in the room can be rotated to generate different musical sounds.
All interactions occur inside the browser, and Underdog says the model runs without requiring an online connection. The company also makes the model available for anyone to download.
How far can a model this small go?
Because of its small footprint, Woof is unlikely to replace larger cloud models, but it may be sufficient for local assistant experiences or for producing personalized audio like podcasts directly on a user’s device. The demo follows Underdog’s earlier release of Bark, a 167 MB speech‑to‑text model aimed at edge devices for transcription and dictation.
Company goals and other releases
Sigil Wen, founder of Underdog, told TechCrunch he began building the app because he preferred developing his own system that runs on users’ devices rather than relying on third‑party note takers like Granola. He said the company aims to train strong, American‑made open models that let people have private AI, referring to the company’s manifesto.
In addition to Woof and Bark, Underdog has released a 27‑billion‑parameter model intended to run on laptops, and it is offering a waitlisted app positioned as a locally hosted alternative to services such as Muse and Instinct.
Release timing
Underdog announced the Woof demo and related materials on Twitter on November 5, 2026.
