A speech model small enough to fit in a browser tab is the latest release from Underdog, an AI startup focused on private, on-device models. The company released a browser game built with its new voice model, called Woof, which TechCrunch describes as only 32MB, according to TechCrunch and Mezha.
The model. Woof has about 31 million parameters and, according to Underdog’s own post, a footprint of 31.7MB, so the 32MB figure is a rounding. Underdog says it can run on most devices without an internet connection, and that it is open for anyone to download. TechCrunch’s article does not give benchmarks, the licence, or which devices were tested.
The game. The demo is set in a retro room with a typewriter, a CRT monitor and a desk lamp. Typed text is spoken back, and a tuner generates music. The intro warns about intrusive cloud providers, which fits the company’s message that voice AI can run locally. The post invites players to defeat a character called Darkcloud for an invite.
The company’s other releases. Underdog launched Bark weeks earlier, a 167MB speech-to-text model for transcription and dictation. It has also released a 27-billion-parameter model that runs on laptops, according to TechCrunch. The company also has a waitlisted app that it positions as a locally hosted alternative to AI agents such as Meta’s Muse and Instinct.
What the founder says. Sigil Wen, Underdog’s founder, told TechCrunch that the aim is to allow people to have private AI. Wen began building the app because of a wish for a personal system instead of cloud notetakers such as Granola.
The limits. Other coverage notes that a model this small is unlikely to replace larger voice systems, though it could power local assistants or personalized audio. One commentary points out that Underdog does not say which devices were tested or what the memory ceiling is, and no independent testing is reported yet.
Why it matters. Most voice assistants send audio to a remote server. A model that runs on the device keeps the audio local, works offline, and costs the developer nothing per request. For people and companies in the US and UK that worry about privacy, small models like this show that useful speech features can move closer to the user, even if quality is still to be proven.
What to watch. Look for independent quality tests, the licence terms, the launch of Underdog’s app, and whether larger companies release similar on-device models.
Sources: TechCrunch and Mezha. Details are as reported by those outlets and Underdog’s posts.
Studio microphone with pop shield (representative image), Wikimedia Commons, CC BY-SA 3.0
