Models · 2026-11-06

Underdog shows off a 32 MB speech model that generates voice and music

Startup Underdog released Woof, a tiny open speech model that runs in a browser without internet, and showed it off in a game.

✨ What happened

Most new AI models need powerful servers because they are very large. Startup Underdog took a different route and built a small model that runs directly on your phone or computer. The model is called Woof and weighs only 31.7 MB.

To demonstrate it, the company made a small game set in a retro-style room with a typewriter, a CRT monitor and a desk lamp. You can type on the keyboard and the game speaks your text back. You can also turn a tuner to generate different music. Everything runs in the browser, and players can defeat the character Darkcloud to get an invite.

📌 Why it matters

The model has 31 million parameters and is open, so anyone can download it. According to the company, it can run on most devices without an internet connection. A model this small will not replace large AI systems, but it may be enough for simple tasks, such as a local voice assistant or personalized podcasts.

Underdog previously released Bark, a 167 MB speech-to-text model also designed to run on devices. The company also has a 27-billion-parameter model that can run on laptops, and a waitlisted app that it positions as a locally hosted alternative to Muse and Instinct.

🔹 What the founder says

Underdog founder Sigil Wen told TechCrunch he started building the app because he wanted to avoid notetakers like Granola and build his own system, with AI models that run on users' devices and let them own their data. He believes the most capable models of today will run on existing devices within months. The company published his manifesto on its website.

Sources: 🔗 TechCrunch — AI🔗 DOU/dev.ua