The startup Underdog launched on November 6, 2026 Woof, a voice synthesis model with 31 million parameters and about 32 MB, that runs entirely in the browser, with no internet connection. To demonstrate it, the company published a browser game with a retro office theme, where the user types on a virtual typewriter and hears the text spoken back.
Underdog launches 32 MB voice model in the browser
Startup Underdog launched Woof, a 32 MB voice model that runs in the browser without internet and is open for download.
Marcelo Santiago
Criador do bridgecode
The biggest AI newsletter in the World
Get AI articles in your inbox
Technical curation on models, agents and coding tools. No spam, unsubscribe anytime.
In this article you will see:
- What Woof is and how it works
- What the demo shows
- How it compares to other voice models
- What has not been published yet
- Why this matters for people building with AI
What is Woof
Woof is a text to speech model with 31 million parameters, compressed to about 31.7 MB. Underdog says it runs on most devices without an internet connection and is open for download by anyone.
The demo is a browser game: an office with a typewriter, CRT monitor and lamp. The user types on the keyboard and the model generates speech in real time, locally. It is also possible to turn a tuner to generate music. Everything happens in the browser, without sending data to external servers.
The startup was founded by Sigil Wen, a fellow of the Thiel Foundation, under Conway Research. It launched an invite only beta in October for Mac and Windows, with versions for Linux, iPhone and Android promised. Investors include Andreessen Horowitz, Khosla Ventures and the Anthology Fund of Menlo Ventures and Anthropic.
What the demo shows
The model runs in the browser because it is small enough to be downloaded and run locally. Underdog says it works on most devices without internet. The model is open for download, which allows developers to integrate it into their own applications.
Underdog had already launched Bark, a speech to text model of 167 MB, also designed to run on edge devices for transcription and dictation. The company's main assistant model has 27 billion parameters, fine tuned from Qwen3.8-27B, running on its own inference engine.
The company's thesis is that the assistant runs entirely on device, keeping user data locally, with no ads or data mining. Wen told TechCrunch that he started building the app because, instead of using notetakers like Granola, he wanted to build his own system with AI models that run on users' devices.
How it compares to other voice models
At 32 MB, Woof is small enough to be embedded in an app or web page without a long download. This makes local voice practical on older phones and cheap laptops, not just on high end hardware. Many popular open voice models are several times larger, and most commercial voices still run in a data center.
Underdog says that in a few months it will be possible to run today's most capable models on the devices people already own.
What has not been published yet
Several important details have not been disclosed. Underdog has not published license terms for Woof, a list of supported languages or voices, or quality benchmarks. The TechCrunch report also does not include this data.
A 31 million parameter model will have limitations and will probably not reach the expressiveness of large cloud based voices. The compact size brings quality trade offs that the company has not quantified publicly.
Why this matters for AI builders
Voice models that run locally, without internet, open up possibilities for apps that need privacy, low latency or offline operation. Voice assistants on devices, accessibility tools, games and embedded interfaces can use voice without relying on external servers.
Woof shows that viable voice synthesis can fit in tens of megabytes. For people building AI products, it is worth following the evolution of these compact models, especially once the license and benchmarks are published.
Frequently asked questions about Underdog's Woof voice model
What is Woof?
It is a text to speech voice synthesis model with 31 million parameters and about 32 MB, launched by startup Underdog.
Does Woof run offline?
Yes. It runs entirely in the browser, with no internet connection.
Is the model open?
Yes, it is open for download by anyone.
What are the limitations?
Underdog has not published a license, a list of supported languages or voices, or quality benchmarks. Quality should be lower than large cloud voices.
What is a voice model this small for?
For offline apps, privacy, low latency and embedded devices, where cloud models are not viable.
Use cases for developers
For people building AI products, a 32 MB local voice model opens doors that cloud models do not. Apps that need to work offline, such as field tools, embedded devices and apps for regions with unstable connectivity, can add voice without depending on external servers.
Privacy is another point. When synthesis happens on device, user text never leaves the device. This matters for health, finance and communication apps where sensitive data should not travel over the network.
Latency also drops. With no round trip to a server, voice response is immediate. For conversational interfaces, that changes the feel of fluidity.
Woof still has no license or published benchmarks. Until Underdog publishes those details, production use requires caution. But as a proof of concept that viable voice synthesis can fit in tens of megabytes, the launch is a sign of what is coming.
For deeper reading, see also how to run models locally with TensorRT on DGX Spark and tips for validating integrations with Transformers.
Source: TechCrunch
PY
Tags:#radar#ia
Developer and creator of bridgecode, an agentic development environment for orchestrating squads of AI agents. Helps people going from vibe coding to professional development move from idea to shipped product.
Related articles
Newsletter
Rich Sutton reinforces that compute alone does not create intelligence
Turing Award winner Rich Sutton reinforces that raw compute is not enough for intelligence and calls for new methods, continuous learning and world models.
5 min
Newsletter
AI auditors detect misbehavior but generate false alarms
AI models catch a good portion of other agents' misbehavior, but they generate a lot of false alarms. See the numbers from Andy Hall's study.
6 min
Newsletter
Palo Alto Networks in talks to acquire agent security startup
Palo Alto Networks is in talks to acquire Clutch Security, an AI agent credential security startup, for US$500 million.
5 min