# Underdog 32MB speech model runs in a browser, offline

> Underdog's Woof text-to-speech model is about 32MB, runs in the browser and is open to download. Here is what founders should know and test this week.

- URL: https://tools.launchllama.co/blog/underdog-32mb-speech-model-woof-browser-demo
- Published: 2026-11-07
- Publisher: Launch Llama (https://tools.launchllama.co)

TL;DR
- Underdog released Woof, a text-to-speech model of 31.7 MB (31 million parameters) that runs in the browser. The company says it works on most devices without an internet connection.
- The model is open for anyone to download. Underdog showed it in a browser game where a typewriter speaks back whatever you type.
- Weeks earlier Underdog released Bark, a 167MB speech-to-text model for edge devices. Together they cover transcription and speech generation locally.

## What happened

Underdog, a startup building small models that run on your own devices, released a game to show off its new text-to-speech model, Woof. The model is 31.7 MB, has 31 million parameters, and runs inside the browser. Underdog says it can run on most devices with no internet connection, and that it is open for everyone to download.

The game is set in a retro room with a typewriter, a CRT monitor, a desk lamp and a music tuner. You type on the typewriter with your keyboard and the game speaks the text back. Turning the tuner generates different kinds of music. Underdog says all the voices in the game are Woof. The intro also takes a jab at "intrusive" cloud providers.

This comes weeks after Underdog launched Bark, a 167MB speech-to-text model for transcription and dictation on edge devices. Underdog has also released a 27-billion-parameter model that can run on laptops, plus a waitlisted app it positions as a locally hosted alternative to Muse and Instinct. Founder Sigil Wen says the goal is to train capable, open, American-made models that give people private AI.

## What this means for founders

- **Cost:** a model that runs on the user's device means no per-minute cloud speech bill for that feature. The source does not give any pricing, so check the license before you build on it.
- **Privacy:** if the audio never leaves the device, you can offer a simple private-by-default story. Wen's stated goal is letting users own their data.
- **Good fit:** TechCrunch says a model this small likely won't replace bigger ones, but it may be enough for local assistant experiences or personalized podcasts.
- **Unknowns:** the source gives no voice quality benchmarks, language support, latency numbers or license terms. Treat it as a demo until you have tested it yourself.
- **Pairing:** with Bark (speech-to-text, 167MB) and Woof (text-to-speech, about 32MB), a fully local voice loop looks possible. The source does not say Underdog has combined them.

## How to try it today

- Open Underdog AI's post on X (@UnderdogAI, November 5, 2026) and follow the link to the game.
- Type a few lines on the typewriter and listen to how the voice handles names, numbers and longer sentences.
- Turn off your internet connection after the page loads and test again, to check the offline claim for your own device.
- Defeat Darkcloud in the game if you want an invite, per Underdog's post. What the invite unlocks is not confirmed in the source.
- Find the model download through Underdog's channels. The source says it is open to download but does not give a link or license terms, so read the license before shipping anything.

A quick test script to run through the typewriter:

Order 1042 ships on March 3rd. Call Dr. Nguyen at 555-0143.

## Related tools on Launch Llama

Tools from the Launch Llama directory that fit with this news:

ToolWhat it doesWhy it matters hereDesert Ant LabsFor developers: Desert Ant Labs provides small AI models for speech, text, and vision that run offline on devices so you ship private, fast features.Desert Ant Labs also provides small speech, text and vision models that run offline on devices, so it is the closest comparison for on-device voice features.Smallest.aiFor voice agents: Smallest.ai's Lightning V3 delivers 100ms latency text-to-speech with instant voice cloning, supporting 20+ languages.Smallest.ai offers cloud text-to-speech with 100ms latency and voice cloning, a useful benchmark if you are weighing local against hosted speech.SpeechGen AIFor content producers: SpeechGen AI converts text to natural speech with voice cloning in multiple languages so you create audio content in minutes.SpeechGen AI converts text to speech with voice cloning in multiple languages, an option if you need a ready-made hosted tool rather than a model to self-run.

[Browse all tools](https://tools.launchllama.co/tools) or [list your own tool](https://tools.launchllama.co/submit).

## Llama's take

I would try this now if you are building a voice feature where privacy or offline use matters, or if cloud speech costs are starting to hurt. The demo takes a couple of minutes, and the model is open to download, so testing is cheap.

I would wait if you need proven voice quality, many languages, or clear commercial terms. The source has none of that yet, and a game demo is not a production test.

The thing to watch is whether Underdog publishes license terms and quality comparisons for Woof and Bark. That decides whether this is a toy or a real option for shipping private voice features.

## FAQ

### What is Underdog's Woof model?

Woof is a text-to-speech model from Underdog AI. It is 31.7 MB with 31 million parameters, and Underdog says it can run on most devices without an internet connection. The model is open for anyone to download.

### Can the Underdog 32MB speech model run in a browser?

Yes. Underdog built a game around Woof that runs in the browser and speaks back whatever you type on a typewriter. All voices in the game are generated by Woof.

### Is Woof a replacement for larger speech models?

Not likely, according to TechCrunch. It may be enough for local assistant experiences or personalized podcasts. Benchmarks, language support and license terms are not confirmed in the source.

## Sources

- [Underdog shows off a 32MB speech model in new demo](https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo) (TechCrunch AI)

---

More for agents: https://tools.launchllama.co/llms.txt · https://tools.launchllama.co/catalog.md · https://tools.launchllama.co/mcp
