Underdog's Woof packs a 32MB speech model into a browser game

What is Underdog's Woof model?

Underdog, the AI startup building open models that run on user devices, has released a new voice model called Woof that is only 32MB large, the company and TechCrunch reported in early November. The model has 31 million parameters and is open for everyone to download.

TechCrunch reported that the model can run on most devices without any internet connection, with speech synthesis calculated locally in the browser. All of this works through a game Underdog built for the demo, which features a room with a retro tech theme, a typewriter and a CRT monitor.

The demo's intro warns users about "intrusive" cloud providers in the town - an apparent jab at cloud-based models - and players can type anything on the typewriter and the game speaks it back. A tuner in the game generates different kinds of music, all with Woof voices.

Woof sits in a growing family of small open models on the site's radar. Underdog previously released Saluki 27B, a compressed laptop-runnable model, and the company says it now also has a 27-billion-parameter model designed to run on laptops.

Why do small speech models matter?

Most frontier AI companies are racing to make bigger and better models, but all of them need massive amounts of processing power, TechCrunch noted. On the other end, some researchers and startups are creating smaller models that run on edge devices like phones and computers.

While Woof is unlikely to replace bigger models, TechCrunch noted it may be enough to drive local assistant experiences or create personalized podcasts. For privacy-focused use cases, an on-device speech model means the text and audio never leave the user's machine.

Sigil Wen, Underdog's founder, told TechCrunch he started building the app because he wanted his own system with AI models that run on users' devices and let them own their data. "I think we will be able to run most capable models of today on the devices you already own within months. 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," he said.

The trade-off is real: a 32MB model cannot match the fidelity of large cloud text-to-speech systems, but it can work in places with no connectivity, at zero marginal compute cost, and without streaming audio to a provider. Woof joins a wave of open-weight models built for phones and laptops, like Liquid AI's LFM3 family, and it is not the only open speech model in the category: NVIDIA's Nemotron 3.5 Transcribe targets similar edge transcription use cases.

What else has Underdog released?

Woof's demo comes weeks after Underdog launched Bark, a 167MB speech-to-text model also designed to run on edge devices for transcription and dictation, TechCrunch reported.

In addition to the two voice models, Underdog has released a 27-billion-parameter model that can run on laptops, along with a waitlisted app that the company positions as a locally hosted alternative to tools like Muse and Instinct, according to TechCrunch.

The developers have also put Woof on Hugging Face so anyone can download it and experiment, dev.ua reported, noting the license only allows non-commercial use without separate permission.

What is the practical takeaway?

Woof is a demonstration that speech generation is no longer confined to the datacenter. The model's download, open weights and browser-ready deployment mean developers can build voice features into products that work offline and keep user data local.

The limits matter too: 31 million parameters is roughly three orders of magnitude smaller than frontier models, so output quality, accent coverage and long-form audio will fall short of cloud systems. Underdog's own website positions the model as part of a private-AI line rather than a replacement for big cloud voices.

For buyers and builders, the pattern to watch is whether tiny open speech models become good enough for assistant and dictation defaults, the same way small open LLMs have carved out a laptop-and-phone niche.

How does Woof compare with other small speech models?

Woof is not the only small open speech model on the market. NVIDIA's Nemotron 3.5 Transcribe, released in late October, is an open-weight speech model that transcribes 40 languages and targets similar edge use cases, while Liquid AI's LFM3 family is built specifically for phones and laptops.

The differentiator Underdog claims for Woof is size: at 32MB it is roughly one-fifth the footprint of its own text-to-speech predecessor work and can run in a plain browser tab, which removes the need for a native app or a runtime install. Whether that is enough to make it a default for developers building voice interfaces remains to be seen; the gate for adoption is usually consistency and latency more than raw parameter count.

A 32MB model also fits the broader push toward tiny open models for private AI. Underdog's founder frames the company's mission as training open, capable models that let people own their data, and the Woof demo is deliberately presented as a local alternative to cloud voice services.

What should developers watch next?

The two practical questions are licensing and quality. The model's Hugging Face release is limited to non-commercial use without separate permission, which means commercial products need to check the license terms before shipping, dev.ua reported.

On quality, the realistic test is a side-by-side with cloud text-to-speech at the same task: short assistant replies are likely to be usable, while long-form narration and multi-speaker audio are where small models typically fall apart.

If Underdog's prediction holds, the bigger shift is strategic rather than technical. Running today's most capable models on devices people already own would change where the cost of inference lives and who controls the data, and speech is one of the first modalities where a tiny open model can credibly demonstrate that path.