Underdog shows 32MB speech model running in a browser
Startup Underdog released a browser game powered by Woof, a 31-million-parameter text-to-speech model of about 32MB that runs offline on most devices, according to TechCrunch.
While many AI companies race to build ever larger models that need vast amounts of computing power, a few researchers and startups are going the other way. One of them, Underdog, has released a browser game built on a new voice model called Woof that is roughly 32MB in size, TechCrunch reported.
A game as a demo
The game is set in a retro-themed room with a typewriter, a CRT monitor and a desk lamp that clicks when switched on. Its intro warns about "intrusive" cloud providers in town, which TechCrunch read as a jab at cloud-based models. Players can type text on the typewriter and hear it spoken back, or turn a music tuner to generate different kinds of music.
In a post on X on 5 November, Underdog said: "We made a game for our new frontier text-to-speech model, Woof. 31.7 MB, so small that it runs in your browser." The post added that all voices in the game are Woof and that defeating a character called "Darkcloud" earns an invite.
About the model
Underdog says Woof runs on most devices without an internet connection and works inside a browser. It has 31 million parameters and is available for anyone to download. TechCrunch noted that such a small model is unlikely to replace larger ones but could power local assistant experiences or personalised podcasts.
The release follows other launches from the company. Weeks earlier it released Bark, a 167MB speech-to-text model for edge devices aimed at transcription and dictation. It has also released a 27-billion-parameter model that can run on laptops, and it has a waitlisted app that it positions as a locally hosted alternative to Muse and Instinct.
The founder's pitch
Founder Sigil Wen told TechCrunch he began building the app because, rather than using notetakers such as Granola, he wanted his own system, with AI models running on users' devices so that people own their data. He said he expects most of today's capable models to run on devices people already own within months, and that Underdog wants to train the best open, American-made models to enable private AI.
Why it matters
Models that run locally avoid sending data to the cloud and can work without a connection, which is the core of the privacy argument Underdog is making. Whether tiny models can match the quality of large hosted ones is a separate question, and the demo is a showcase rather than a benchmark. The claims about size, parameters and capabilities here come from Underdog and TechCrunch's description of the demo, and were not independently tested.
Sources
Written with automated tools from the sources above and published without editor review. Spotted an error? Tell us · Corrections Policy
Related stories
California sues OpenAI over rogue AI agent incidents
California's attorney general has sued OpenAI, alleging that its model testing practices lacked adequate safeguards after a July incident involving Hugging Face. OpenAI says the suit singles it out over an industry-wide challenge.
What Huang's Nvidia exit could mean for open-source AI
Jensen Huang's resignation as Nvidia chief removes one of the most prominent backers of open-weight AI models. Analysts quoted by Scientific American and TechCrunch see risks for talent, policy and a pending Hugging Face deal.
Anthropic adds Sheets, Whiteboards and Data to Claude
Anthropic launched collaborative spreadsheet and whiteboard tools in beta and made Claude Data generally available. Reports differ on when Claude Meetings arrives.
New York opens registration under its AI safety law
New York has opened a portal for large frontier AI developers such as OpenAI and Anthropic to register by January 1, 2027 under its RAISE Act.
