📅 November 6, 2026 • TechCrunch
Underdog demos Woof, a 32MB speech model that runs fully offline in the browser
Per TechCrunch, the company Underdog has shown off Woof, a speech model that weighs just 32MB and packs 31 million parameters. The company says the model runs on-device and fully offline - no internet connection is involved at any point in the loop - and its weights are open for download, so developers can pull the model and run it locally inside their own apps and pages. Everything the demo does happens on the machine in front of the user, a strikingly compact footprint for speech synthesis, a job that has mostly meant a datacenter round trip so far. A download this small changes the product shape of voice entirely - it fits where server-side speech never could.
The size is the story here. Voice interfaces usually lean on cloud APIs, which means every utterance travels to a remote server, comes back with added latency, costs money on every call and routes user audio through someone else's infrastructure. A 32MB model that speaks from inside the browser flips that pattern. The audio never leaves the device, inference carries no per-call cost, and the feature keeps working when the connection drops. In a category where responsiveness is the whole product experience, moving synthesis to the client is a genuine shift in engineering - the voice layer stops being a network service and becomes an asset you ship together with the app, versioned and cached like any other resource.
The release also fits a wider pattern at Underdog. Weeks earlier, per TechCrunch, the company shipped Bark, a 167MB speech-to-text model built for edge transcription. And Underdog has also shipped a 27-billion-parameter model that runs on laptops. Put the three releases together and a consistent thesis emerges - useful AI does not have to live in a datacenter, small open weights can sit directly on the device next to the user, and one lab can cover listening, speaking and heavier reasoning across browsers and laptops alike. The common thread is packaging - open weights at sizes a hobbyist can download and run the same afternoon. There is also a distribution angle - weights this small travel inside the app itself and reach users without special infrastructure.
IMPACT: For builders, a downloadable 32MB voice model means speech features can ship inside web apps without a cloud inference backend - lower latency, no per-call API spend, and audio that stays local by default, which quietly simplifies the privacy conversation too. Combined with Bark for on-device transcription, a small team can cover the full voice loop offline on open weights, and voice UX turns from an integration project into a client-side feature that ships with the frontend. It also lowers the bar for experiments - a voice assistant in a single HTML file starts to look plausible.
Sources: techcrunch.com, mezha.net, tech.yahoo.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
