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

Underdog has released a playable browser demo built on Woof, a speech generation model that is just 32MB — small enough to run on most devices with no internet connection at all.

The demo takes the form of a game. You land in a retro-tech room containing a typewriter, a CRT monitor, and a desk lamp that clicks audibly when you flip its switch. The intro warns players about "intrusive" cloud providers in the town — a not-so-subtle jab at cloud-based AI models.

The interaction is the point. Type anything on the typewriter with your keyboard and the game speaks it back to you. Rotate the music tuner and it generates different kinds of music. All of it runs locally, inside the browser, powered by Woof.

The model has 31 million parameters, and Underdog has made it open for anyone to download. To put that in perspective: frontier language models run to hundreds of billions of parameters. Woof is roughly the size of a single high-resolution photo folder — a rounding error by today's standards.

What does this change for users and developers?

Don't expect a 32MB model to replace the big cloud systems. It won't. But Underdog's bet is that it doesn't need to. A model this small could still drive local assistant experiences or generate personalized podcasts entirely on-device — no API bills, no latency, no data leaving the phone.

For developers, the signal is that the edge-AI toolchain is maturing fast. You can now download an open speech model, ship it inside a browser game, and have it speak user input offline. That was a server-side job until quite recently.

A growing local-first stack

Woof isn't Underdog's first move on edge devices. Weeks earlier, the company launched Bark, a 167MB speech-to-text model also built to run on-device for transcription and dictation. A useful frame: Bark is about five times the size of Woof, and both fit comfortably on a phone.

The startup has also released a 27-billion-parameter model that runs on laptops, plus a waitlisted app it positions as a locally-hosted alternative to Muse and Instinct.

Founder Sigil Wen told TechCrunch he started building the app out of personal frustration. Instead of using notetakers like Granola, he wanted his own system powered by AI models that run on users' devices — and let users own 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. "We just want to train the best open models that are American-made, that are open, that are capable and allow people to have private AI." He was referring to his company manifesto.

The stakes in the small-model race

Most AI labs are racing to build bigger, hungrier models that demand massive processing power. Underdog sits at the other end of that spectrum, betting that private, offline, on-device AI is a category enterprises and consumers will actually pay for.

That's months, in Wen's telling, not years — today's capable models running on hardware you already own. If that timeline holds even partially, the practical question for anyone building voice or transcription features shifts from "which cloud vendor" to "which open model we can ship inside the product." Woof's 32MB demo is an early, tangible answer.