# Underdog's 32MB voice model runs in a browser with no internet connection

> The 31-million-parameter speech model powers a retro demo game and is open to download. The startup also makes on-device transcription and a 27B laptop model.

- URL: https://devisdead.ai/news/underdog-s-32mb-voice-model-runs-in-a-browser-with-no-internet-connection
- Published: 2026-11-08T23:31:14Z
- Author: DevIsDead Newsroom
- Section: Dev
- Tags: underdog, on-device ai, speech, open models
- Originally reported by TechCrunch (https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo/).

While the big labs build ever-larger models, startup Underdog is going the other way. It has released a demo game powered by its new speech model, Woof, which is just 32MB, Ivan Mehta reports for [TechCrunch](https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo/).

## The demo

The game is a retro-styled room with a typewriter, a CRT monitor and a desk lamp that clicks when you flip it. Type anything on the typewriter and the game reads it aloud; turn the music tuner to generate different kinds of music. An intro pokes fun at "intrusive" cloud providers in town.

It all runs in the browser on Underdog's model, which the company says works on most devices with no internet connection. Woof has 31 million parameters and is free to download.

## The bigger plan

A model that small won't replace big cloud voices, but it could drive local assistants or personalized podcasts. A few weeks earlier it shipped Bark, a 167MB model that turns speech into text on the device for dictation and transcription. There's also a 27B model sized for laptops, and has a waitlisted app pitched as a local alternative to Muse and Instinct.

Founder Sigil Wen says he started because he wanted his own note-taking system on models that run on his devices, so users own their data. He thinks most of today's capable models will run on devices people already own within months, and wants to train the best open, American-made models for private AI.

## Why developers should care

Tiny, offline speech models change the economics of voice features: no per-request cost, no latency to a server, and no audio leaving the device. For apps with privacy-sensitive users or spotty connectivity, a 32MB model you can bundle is worth trying before you sign another API contract.

**[Scoreboard](/scoreboard) update:** AI +1. A capable voice model shrinks to 32MB and runs offline in a browser.

Stay undead.