Building ever-larger AI models remains the dominant trend—but another parallel trend is underway: shrinking models down to fit inside your browser. On November 6, startup Underdog released a speech model Woofweighing just 31.7 MB and containing 31 million parameters—open-source, downloadable, and fully functional offline.
Instead of publishing a technical blog, they built a mini-game to demonstrate it: a retro-tech room with a typewriter, a CRT monitor, and a lamp that chimes when pressed. Type anything on the typewriter, and the game voices it aloud in real time using Woof; twist the tone knob to generate music in different styles—all running entirely in-browser, no internet required.
What does 32 MB mean?
Compare it to familiar benchmarks: a one-minute 4K video shot on a smartphone is ~400 MB; WeChat’s installer often exceeds several hundred MB. Woof’s entire model is 31.7 MB—barely larger than a single smartphone photo—yet it performs text-to-speech: converting any text into human-like speech.
One detail in the game stands out: at launch, players are warned to watch out for the town’s “nosy” cloud provider. It’s a direct jab at cloud-based models—Underdog founder Sigil Wen has stated he refuses to use note-taking tools like Granola and instead wants a locally run, user-owned system.
Their roadmap: small models + edge deployment + open source
- WoofA 31.7 MB, 31-million-parameter text-to-speech model, open-source and downloadable, with local inference in the browser.
- BarkA 167 MB speech-to-text model released weeks ago, designed for edge-based transcription and dictation.
- 27B-parameter modelOfficially stated to run on laptops (a vendor claim; actual performance depends on memory and quantization method).
- Local app versionCurrently in waitlist phase, positioned as a local alternative to cloud-based agents like Muse and Instinct.
Sigil Wen publicly stated, “Within months, we’ll be able to run most of today’s capable models on the devices you already own,” and emphasized building “U.S.-made, open, private AI models that users truly own.” Commercially, this is a differentiation strategy; technically, it bets on one thing: the pace of progress in edge-side compute power and model compression is faster than most expect.
What small models can and cannot do
Frankly, a 32 MB TTS model cannot replace ElevenLabs commercial voice services from Alibaba or ByteDance. Its voice timbre richness, emotional expressiveness, and multilingual coverage are limited, and stability on long texts requires real-world testing.
Yet its use cases are highly specific—and happen to align precisely with pain points of cloud-based solutions:
- Offline device narrationIn-vehicle systems, toys, industrial terminals, and voice prompts in offline environments—no data connectivity required and no risk of network interruption.
- Privacy-sensitive scenariosText narration for healthcare, legal, or internal meetings—content never leaves the device.
- Zero marginal costLocal inference incurs no per-character fees; cost structures differ entirely for personal podcasts, audiobook drafts, and other long-text applications.
- Controllable latencyWithout network round-trips, first-token latency can be nearly instantaneous.
Conversely, commercial voice-over requiring “human-like主播” quality, multi-character audio dramas, and emotionally expressive scenarios still rely on cloud-based large models. This contrasts with what we saw in NVIDIA Nemotron 3.5 Transcribe benchmark — there, a 34-billion-parameter model supports 40 languages and processes one-hour audio on a single GPU, pursuing “bigger and stronger”; Underdog pursues “smaller and closer.”
Another category of experimentation exists in the speech domain:Block’s Dessa builds full-duplex dialogue using speech-only training,as well as Industry discussions on “speech AI hasn’t yet reached its ChatGPT moment”Consensus holds that speech interaction hasn’t taken off—but disagreement persists over whether breakout will hinge on cloud intelligence or on-device immediacy.
A notable signal: open-source small models are turning “American-made”
Over the past two years, representative on-device open-source small models have largely originated from Chinese teams—Qwen’s compact series, GLM’s Flash variants,DeepSeek and distilled models. U.S. startups like Underdog now explicitly market “American-made, open” as a key selling point, indicating open-source small models have also become part of geopolitical competition.
Practical impact for developers: more choices. Previously, on-device speech meant either using system-built-in TTS (fixed voice, no customization) or training a small model yourself (high cost). Now you can directly download a 32MB weight file and embed it into an app or webpage to test—near-zero trial-and-error cost.
Frequently Asked Questions (FAQ)
Is Underdog Woof free?
Model weights are open-source and freely downloadable; the game demo runs directly in-browser. Underdog’s local app is currently invite-only; its business model isn’t fully disclosed yet but will likely follow a “open model + app subscription” approach.
Is 32MB sufficient for voice quality?
Sufficient for notification sounds, text-to-speech reading, and offline assistants—but insufficient for commercial voice-over. Small models’ main limitations lie in emotional expressiveness, voice diversity, and long-text stability; cloud-based large models remain necessary for multilingual, broadcast-grade performance.
What hardware specs are required for local inference?
The 31-million-parameter model imposes extremely low hardware requirements—runs on ordinary laptops or even mobile browsers, no discrete GPU needed. The official game demo runs pure browser-based inference, confirming CPU inference is viable.
Can it be used directly in commercial projects?
Depends on the specific open-source license (Underdog hasn’t consistently clarified this across all channels—always verify the LICENSE file in the repository before use). Additionally, speech synthesis involving voice cloning and personality rights requires authorization when using real human voices as references.
Summarize
Underdog’s Woof is not about beating cloud-based TTS—it delivers value by lowering the barrier to “speech generation” to a single download. When models shrink enough to fit inside web browsers and earbud cases, “internet connectivity” ceases to be a prerequisite for product design.
Want to discover more useful AI tools? Check out our AI Model Library and Tool Comparison Engineor continue reading:Nemotron 3.5 Transcribe Review · Document your initial hypothesis and metric selection with a timestamp. Experiments show this does not fully prevent influence, but at least lets you see afterward what changed in your thinking · Full In-Depth Evaluations.