Underdog's Woof: A 32MB Text-to-Speech Model That Runs Offline
Woof is a text-to-speech model from Underdog, the on-device AI startup founded by Sigil Wen. It has 31 million parameters, takes up about 32MB, and runs inside a web browser; the company says it works on most devices without an internet connection and is open for anyone to download. License terms, languages, and benchmarks have not been published.
On this page
What Underdog released
Underdog, the startup building a private AI assistant that runs on your own hardware, has shown off Woof, a text-to-speech model with 31 million parameters and a download of about 32MB. The company says Woof runs on most devices without an internet connection and is open for anyone to download. For comparison, many popular open voice models are several times larger, and most commercial voices still run in a data center.
The demo is a browser game set in a retro room with a typewriter, a CRT monitor, and a desk lamp. You type on the typewriter and the game speaks your text aloud, and a tuner dial produces different kinds of music, though it is not clear whether that sound comes from Woof itself. The intro jokes about "intrusive" cloud providers, which tells you where the company is aiming.
Who is behind it
Underdog was founded by Sigil Wen, a Thiel Fellow, under a company called Conway Research. It launched as an invite-only beta in October for Mac and Windows, with Linux, iPhone, and Android versions promised. Its backers include Andreessen Horowitz, Khosla Ventures, and the Anthology Fund run by Menlo Ventures and Anthropic.
The pitch is that the assistant runs entirely on the device, so personal data stays there, and that the product will never be ad-supported or mine user data. The main assistant model is a 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B, running on an in-house inference engine. Wen told TechCrunch he expects to "be able to run most capable models of today on the devices you already own within months".
Woof joins Bark, a speech-to-text model for local transcription and dictation released a few weeks earlier. Bark 3.0 was built from NVIDIA's Parakeet, and its weights were published under a CC-BY 4.0 license; its own benchmark put it level with Parakeet on average rather than ahead.
Why a tiny voice model matters
Speech is one of the most personal kinds of data you can hand to a cloud service, and text-to-speech often processes exactly the things you would rather keep private: messages, notes, documents, and reminders read aloud. A model small enough to load inside a browser tab means none of that text has to leave your machine, there is no per-character API bill, and it keeps working on a plane or in a dead zone.
Size also decides where a model can go. At about 32MB, Woof is small enough to ship inside an app or a web page without a long download, which is what makes on-device voice practical on older phones and cheap laptops rather than only on high-end hardware. Paired with an on-device speech-to-text model such as Bark, it points toward a voice assistant loop that never touches a server.
What we still do not know
Several important details are missing. Underdog has not published license terms for Woof, a list of supported languages or voices, or any quality benchmarks, and TechCrunch's report does not include them either. A 31M-parameter model will also have limits, and it is unlikely to match the expressiveness of large cloud voices.
The name can cause confusion, too. Earlier this year Wen described a different Woof: a 4B-parameter assistant model, version 1.1, that handled browser use and tool calls on the device. The new speech model shares the brand but is a separate, far smaller system. Until Underdog publishes a model card with a license and evaluations, Woof is best treated as a promising demo of how small private voice AI can get, not a finished replacement for the voice in your current apps.