---
type: "post"
id: "underdog-releases-woof-a-30-7m-parameter-open-7db0c0"
slug: "underdog-releases-woof-a-30-7m-parameter-open-7db0c0"
title: "Underdog releases Woof, a 30.7M-parameter open-weights text-to-speech model that runs in the browser"
description: "Underdog’s Woof is a 30.7-million-parameter text-to-speech model presented in an interactive browser experience."
retrieval_nugget: "Underdog’s Woof is a 30.7-million-parameter text-to-speech model presented in an interactive browser experience."
published_at: "2026-11-05"
updated_at: "2026-11-05"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["local-models"]
entities: ["Underdog","Woof","Mimi","Kyutai","LibriTTS-R"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Underdog adds a distinct 30.7M-parameter TTS release designed to run locally in a browser via WebGPU, with open weights for noncommercial use."
prior_refs: ["memory:c0eade81-aabd-823f-ace2-779bb658c427","memory:ffeb3a36-956b-81f3-aa88-1a082ffe5626"]
source_url: "https://underdog.ai/escape"
source_urls: ["https://underdog.ai/escape"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:underdog-releases-woof-a-30-7m-parameter-open-7db0c0"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/underdog-releases-woof-a-30-7m-parameter-open-7db0c0/","markdown":"https://newruntime.com/posts/underdog-releases-woof-a-30-7m-parameter-open-7db0c0.md","json":"https://newruntime.com/posts/underdog-releases-woof-a-30-7m-parameter-open-7db0c0.json"}
---

# Underdog releases Woof, a 30.7M-parameter open-weights text-to-speech model that runs in the browser

## Retrieval answer

Underdog’s Woof is a 30.7-million-parameter text-to-speech model presented in an interactive browser experience.

## Summary
Underdog’s Woof is a 30.7-million-parameter text-to-speech model presented in an interactive browser experience. Its small stated download size and claimed CPU speed make local speech generation worth testing, while its non-commercial license sharply limits straightforward product use.

## What happened
Underdog’s page introduces Woof as its text-to-speech model through an interactive experience called “Escape.” The page says Woof is running in the visitor’s browser, on the visitor’s own GPU, using WebGPU. It lists 30.7 million parameters and a 31.7 MB download. The company also states that it runs 18 times faster than real time on a single laptop CPU thread. These are claims on the product page; the supplied material does not include an independent benchmark or enough detail to reproduce the measurement.

The demo embeds Woof in a broader presentation of Underdog’s local generative capabilities. Visitors can hear a voice reading text and see examples of generated music. The page says that all the audio in the experience—every voice and every song—was generated offline by local models. It describes the work as taking place on an M5 Max with 48 GB of memory, and gives example generation times for music clips ranging from 39.3 seconds to 131 seconds. Those examples concern the music demonstrations, not Woof’s speech speed, and should not be treated as TTS benchmarks.

The experience frames on-device execution as the alternative to sending inputs to a cloud service. It says that visitors can click or press Space to continue, reset the voice, and hear text read in “Isla’s voice.” It also presents the model alongside speech-to-text, music generation, and video generation capabilities in the Underdog product. That framing establishes the intended local-first product context, but it does not establish that Woof itself provides each of those other functions. The specific claims tied to Woof are text-to-speech, browser execution through WebGPU, its parameter count, download size, and the stated single-thread CPU speed.

Underdog emphasizes local processing and privacy. Its page says no personal data about the visitor’s device was sent to the company, and that material the experience read about the visitor was read on the visitor’s machine. The site also says the Underdog app offers higher quality and faster performance, but the supplied page does not quantify that difference or explain the app’s hardware requirements. A visitor can explore the browser demo; the page invites users to download Underdog and request an invite, but does not specify a general availability date or price.

The page identifies Woof’s audio decoder as following the design of Mimi, by Kyutai, and says its training data includes LibriTTS-R, credited to Yuma Koizumi and colleagues. It labels the weights “open” for non-commercial use under CC BY-NC 4.0. That is the license stated by the primary page and is a material constraint: the page does not establish permission for commercial deployment. It also provides no language list, voice inventory, model card, or detailed installation instructions in the supplied material.

## What the archive adds
Two recent Field Notes cover distinct compact TTS releases: [Edge0’s Audio8-TTS-Realistic-0.6B](https://newruntime.com/posts/edge0-releases-audio8-tts-realistic-0-6b-an-6d82b2/) and [Silencio-2](https://newruntime.com/posts/silencio-2-18m-parameter-open-weight-english-tts-0e30c8/). Woof is a new release, not a restatement: its differentiating claims here are browser execution, a 31.7 MB download and stated CPU speed. Its non-commercial license also makes it unlike a simple assumption that “open weights” imply general product-use rights.

## Trend signal
Woof offers a limited signal for [generative media becoming application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/): the demo presents speech generation as local software running inside a browser experience rather than as a remote API call. But the supplied material does not show an integration interface, production deployment, or controllable service boundary. This is evidence of a direction to test, not proof that the pattern is established.

## Why it matters
For teams evaluating voice features, a 31.7 MB model and an in-browser WebGPU demonstration suggest a possible route to reduce dependence on cloud speech services in some settings. Local generation may be attractive when data handling, offline operation, or latency matters. The stated CPU speed also makes it relevant to investigate devices without a dedicated GPU, although the page does not provide a reproducible test or compatibility matrix.

The source also separates two deployment questions that teams should keep distinct. A browser demonstration can show that model execution is technically possible on a visitor’s machine, while a customer feature requires predictable performance, a supported voice and language set, and rights to deploy. None of those follow automatically from the demo. The accompanying Underdog app claim may matter for teams considering a native local workflow, but the page gives no comparable quality metric, system requirement, or access detail with which to choose between browser and app.

The license is just as consequential as the performance claim. CC BY-NC 4.0 is expressly non-commercial, so commercial teams should not infer that the weights are usable in a customer-facing product merely because the page calls them open. They would need to review the license and obtain suitable rights if necessary. Teams should also distinguish the demonstrated local experience from the broader product claim: the page does not establish that every device, browser, language, or deployment will perform similarly.

## Limits and uncertainty
The supplied page does not give an announcement date, pricing, general availability terms, or a direct model download path beyond describing the download size. It does not specify supported languages, voice quality measurements, hardware compatibility, memory use, or how the 18-times-real-time CPU figure was measured. Its privacy and offline statements describe this experience, not an independently audited guarantee for every Underdog product. The license is stated as CC BY-NC 4.0, but the page does not explain whether separate commercial terms are available.

The material also does not establish how Woof compares with other compact TTS systems on intelligibility, speaker similarity, latency, or robustness. It identifies design and data influences, but does not provide a model card or enough training-data detail to assess coverage and limitations. The music examples demonstrate other local models in the experience and should not be read as evidence about Woof’s voice quality.

## Reader takeaway
Woof is worth noting as a small TTS model demonstrated locally in a browser, with a company-stated CPU speed claim. The main practical caveat is its CC BY-NC 4.0 license: “open” does not mean automatically cleared for commercial use. Verify language support, device performance and rights before building around it.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)