AI news ·
Underdog releases a 32MB voice model that runs entirely in the browser
Startup Underdog released a browser game using Woof, a 32MB speech model with 31 million parameters that runs entirely on-device without internet.
Startup Underdog has released an interactive browser game built on a speech generation model called Woof that is only 32MB in size. The model, which has 31 million parameters, runs entirely on-device without an internet connection - a direct challenge to the cloud-dependent approach of larger AI companies.
A retro room that talks back
The demo, accessible through a web browser, places users in a room with a retro tech theme: a CRT monitor, a desk lamp with a clickable switch, and a typewriter. An introductory warning takes a jab at "intrusive" cloud providers. Users can type anything on the typewriter using their keyboard, and the game speaks it back through the Woof model. A music tuner in the room can also be rotated to generate different kinds of music.
The model is open for anyone to download. Underdog says it can run on most devices without internet connectivity. While a 32MB model won't replace large-scale speech systems, it points toward locally hosted assistants and personalized audio experiences that keep data on the user's hardware.
Building a private AI stack
The Woof demo arrives weeks after Underdog launched Bark, a 167MB speech-to-text model also designed for edge devices. Founder Sigil Wen told TechCrunch he started the project because he wanted his own AI system rather than relying on cloud-based notetakers. "I think we will be able to run most capable models of today on the devices you already own within months," Wen said. "We just want to train the best open models that are American-made, that are open, that are capable and allow people to have private AI."
Underdog has also released a 27-billion-parameter model that can run on laptops, along with a waitlisted app positioned as a locally hosted alternative to tools like Muse and Instinct. The company's manifesto centers on open models and user-owned data - a stance that diverges from the centralized infrastructure powering most commercial AI services. For professionals exploring on-device AI, Generative AI Courses cover the fundamentals of how models like these are trained and deployed.
Why this matters for creatives and writers
Small, offline-capable models change where and how creative work can happen. A writer drafting in a coffee shop with spotty Wi-Fi, a marketer prototyping voiceover for a campaign on a laptop, or an educator building interactive exercises without student data leaving the device - these scenarios become practical when models shrink to tens of megabytes. The ability to generate speech and music locally, without subscription fees or latency from cloud calls, lowers the barrier for experimentation.
On-device models also sidestep the privacy and ownership questions that come with uploading creative work to third-party servers. For anyone producing original audio content - podcasters, video editors, game developers - local generation means drafts and final assets stay under their control. Underdog's approach suggests that capable speech tools no longer require enterprise-scale infrastructure, which could reshape workflows for independent creators and small teams.