A typewriter, a music tuner and a desk lamp make up the setting for Underdog’s new browser game. Its central feature is Woof, a speech generation model that the company says is small enough to run on most devices without an internet connection.
At 31.7 MB, Woof points to a different path for AI: models designed to work on devices people already own, rather than relying on large amounts of cloud computing.
A game puts Woof to work
The game uses a retro technology themed room as its interface. Players can click the desk lamp’s switch to hear a sound, type words on a typewriter and rotate a music tuner to produce different kinds of music.
When a player enters text, the game speaks it back. Underdog says all of the voices in the experience use Woof, with the model running in the browser. The company’s post describes Woof as a 31.7 MB text-to-speech model that runs in a browser.
The intro also warns players about “intrusive” cloud providers in the town, a pointed reference to cloud based AI. The demo turns that contrast into part of the experience: its speech features work through the small model rather than an internet connection to a cloud service.
What a smaller local model could enable
Woof has 31 million parameters and is available for anyone to download. Its size and ability to run on most devices, according to Underdog, make it an example of edge AI: models that process tasks on a phone or computer instead of depending on remote computing resources.
A compact speech model does not have to compete with larger systems on every task to be useful. The source article suggests that a model like Woof could support local assistant experiences or personalized podcasts. Those uses would make speech generation available in settings where keeping processing on a device is a priority.
Running locally also connects to a broader goal Underdog’s founder described to TechCrunch: building AI systems that let users own their data. Sigil Wen said he began building the app because he wanted a personal system powered by models running on users’ devices, rather than note-taking tools like Granola.
Woof is part of a wider device focused effort
Underdog’s work includes more than text-to-speech. Weeks before the Woof demo, the company launched Bark, a 167 MB speech-to-text model intended for edge devices and tasks such as transcription and dictation.
The company has also released a 27-billion-parameter model that can run on laptops. Separately, it is positioning a waitlisted app as a locally hosted alternative to Muse and Instinct. Together, these releases show that Underdog is pursuing several kinds of AI models and apps that can operate on personal devices.
Wen said he believes capable models could run on devices people already own “within months.” He described the company’s aim as training open models that are capable and support private AI. That is his stated ambition; the Woof game itself demonstrates a narrower capability: text-to-speech and music generation in a browser using a model the company says can work offline.
A practical demonstration, with limits
Woof’s small file size is a notable feature, but the source does not claim it will replace larger AI models. Instead, the demo shows how a focused speech model can make an interactive experience work without depending on a cloud connection.
For users, the distinction is practical. A model that runs on a local device may help keep an experience available without internet access and give people more control over where their data is handled. Underdog’s game offers one early illustration of that approach, while its broader lineup points toward local speech, transcription and other AI use cases.