Voice that doesn't need the cloud: Underdog's 32 MB Woof

In a market where the race is for ever-larger models — GPT-4, Claude Opus, Gemini Pro —, an startup called Underdog is going the other way. Created by Conway Research and founded by Sigil Wen, the company has released Woof, a speech generation model of just 32 MB and 31 million parameters that runs entirely in the browser, without an internet connection. The differentiator is not complexity, but portability.

A game to prove sound fits in a pocket

The way Underdog chose to demonstrate Woof is no accident: a retro browser game called "Escape", set in a room with a typewriter, CRT monitor and a desk lamp that makes sound when toggled. Everything the user types on the typewriter is read aloud by the model. The music tuner generates different kinds of sound. There are no API calls, no intermediary server. The model runs on WebGPU directly on the user's hardware.

The metaphor is not just aesthetic — it's political. The game's introduction warns against "intrusive" cloud providers in the town, a thinly veiled jab at the industry's standard architecture, where every voice command travels to a data center and back with latency, cost and a data log stored somewhere. Underdog is proposing that the entire loop fits in the user's machine.

Why 32 MB matters

The size is the story. Conventional voice interfaces rely on cloud APIs: every utterance travels to a remote server, returns with latency, costs money per call and routes user audio through third-party infrastructure. A 32 MB model that speaks from inside the browser flips that pattern.

For developers, this means voice features can be embedded in web applications without a cloud inference backend — reduced latency, no per-call cost and, implicitly, privacy that comes built in. The model doesn't need WiFi to function. At 32 MB, it is smaller than a modern smartphone photo.

The comparison with leading models is instructive. The top cloud-based speech-to-text and text-to-speech models — from Google, Amazon or OpenAI — are often models with billions of parameters, executed on GPU clusters. Woof has 31 million. It is roughly a thousand times smaller. It does not compete on sound quality or fidelity with larger models. It competes on accessibility and data sovereignty.

Underdog's bottom-up strategy

Woof is part of a sequence of Underdog releases that sketch out a broader pattern. A few weeks before, the company launched Bark, a 167 MB speech-to-text model designed for edge transcription. Additionally, the company released a 27-billion-parameter model that runs on laptops, positioned as a local alternative to Meta Muse and Instinct.

Whether it is sound, transcription or reasoning, Underdog's thesis is consistent: useful intelligence does not need to live in a datacenter. Small open-source models can sit directly on the user's device, with weights that a hobbyist can download and run the same afternoon.

Chip architecture matters too. Underdog developed Husky, an inference engine that moves less data between the computer's CPU and GPU than rival engines. This is significant because the bottleneck is not just model size — it is data movement within hardware. A small model that fits in memory but needs to shuttle gigabytes between chips loses the advantage. Husky solves that.

The business: no subscription, no ads

Sigil Wen, Underdog's founder, is a rare case in Silicon Valley: a self-taught 17-year-old who lived in an AI hacker house with Andrej Karpathy, tested early versions of Claude, Midjourney and Stable Diffusion, and figured out how to run GPT-2 on an Apple Watch. He is a Thiel Fellow with heavyweight investors — a16z led the round, Stripe co-founder Patrick Collison is an angel investor, Khosla Ventures and SV Angel also participated.

But the business model is what stands out most. The app will be free, with no ads and no subscription. Since the processing runs on the user's machine, Underdog does not carry the massive cloud inference costs that weigh on the margins of competitors like Meta or Google. Revenue will come from a tiny percentage fee on payment transactions processed via Stripe, in the style of financial interchange fees. In other words, the company earns when the user spends, not when the data is mined.

Wen says this aligns incentives with the user — the same way a bank or credit card network does. "I don't have to charge you a subscription to run this because my costs are so super low," he told TechCrunch. The contrast with cloud-based agents, which collect user data to train models or sell to advertisers, is intentional.

What comes next

Wen's most honest limitation is clear: a 32 MB model will not replace larger models. But that's not a problem. For local assistant experiences, dictation, personalized podcasting or any app that needs voice without exposing user audio to servers, Woof might be enough.

Underdog has a waitlisted full desktop app and already offers Bark (167 MB) for local transcription and a 27 billion-parameter model for reasoning. The trio covers listening, speaking and reasoning — all running on user devices.

The question is not whether the small model will be as good as the big one. It is whether it will be good enough to solve a problem nobody was solving: private, local, immediate voice on any device, without depending on third-party infrastructure. The answer might be in a retro game with a typewriter and a music tuner — all working without internet.

Sources: TechCrunch, DEV Community, TechCrunch

✓ Independent sources cross-checked and verified before publishing