Underdog demos Woof, a 32MB speech model that runs offline in the browser

A retro room that talks back

Underdog, a startup building compact AI models, has published an interactive demo meant to show how small a useful speech model can be. According to TechCrunch, the demo is a browser-based game set in a retro-themed room containing a typewriter, a CRT monitor and a desk lamp that clicks audibly when you flip its switch. The introduction warns players about "intrusive" cloud providers in the town, which TechCrunch reads as a jab at cloud-hosted AI.

The typewriter is the centerpiece: anything the player types on the keyboard is spoken back aloud by the model. A music tuner in the room can also be rotated to produce different kinds of music. All of the audio generation happens in the browser, and the company says the model can run on most devices without any internet connection.

The model behind it

The voice engine is called Woof. It is a speech generation model that takes up 32MB of storage and has 31 million parameters, TechCrunch reports, and it is open for anyone to download. The demo is essentially a proof of capability: a model small enough to fit comfortably within a phone's storage budget, yet able to turn arbitrary text into speech in real time inside a browser tab.

TechCrunch adds the obvious caveat that a model this small is unlikely to replace large cloud-hosted systems. Even so, it may be enough for local assistant experiences or for generating personalized podcasts, workloads where latency, cost and privacy matter more than peak capability.

Part of a broader portfolio

Woof is not Underdog's first move in this space. Weeks earlier, the company launched Bark, a 167MB speech-to-text model also designed for edge devices and aimed at transcription and dictation. Beyond the two voice models, TechCrunch reports that Underdog has released a 27-billion-parameter model that can run on laptops, and is building a waitlisted app it positions as a locally hosted alternative to Muse and Instinct.

Founder Sigil Wen told TechCrunch that the project started as a personal one: rather than depending on notetaking services like Granola, he wanted a system built on models that run on users' own hardware, so people keep ownership of their data. "I think we will be able to run most capable models of today on the devices you already own within months," Wen said, adding that the goal is to train open, capable, American-made models that let people have private AI.

Why it matters

Most AI progress is measured in scale, but for everyday usefulness the binding constraint is often what can run locally. A 32MB speech model that works offline in a browser signals a different kind of maturity: models tuned to the memory and compute budgets of phones and laptops rather than data centers. For users, on-device speech synthesis means typed text can become audio without that text ever leaving the device, which matters for dictation, accessibility tools and assistants handling sensitive input. For the industry, demos like this are evidence that the gap between cloud capability and edge capability is narrowing, and that pairing open weights with a small footprint is becoming a credible product strategy rather than a research curiosity. Underdog's stated bet is that within months, today's capable models will run on hardware people already own; Woof and Bark are early, tangible steps in that direction.