Underdog is demoing Woof, a text-to-speech model it describes as 30.7M parameters and a 31.7 MB download, running in the browser via WebGPU. The Escape page says open weights for non-commercial use under CC BY-NC 4.0, claims audio on the page was generated offline by local models, and includes an in-browser music tuner. TechCrunch covered the demo on 6 November 2026, quoting founder Sigil Wen on private, American-made open models; TC also notes an earlier Bark STT edge model.
Caveats
License is CC BY-NC 4.0, not Apache. We did not find a confirmed official Hugging Face Woof repository on this pass, so do not claim Hub GA. Speed and offline claims come from the vendor demo and TC; verify on your own hardware before you ship a product on them. This sits beside the earlier invite-only Underdog / Conway private agent story rather than replacing it.
Why this matters
If you care about on-device voice, a ~32 MB WebGPU TTS is a concrete size point you can try without a GPU cluster. Keep the non-commercial license front and center for anything customer-facing, and treat Hub availability as unconfirmed until Underdog publishes an official card. Pair the demo with your own latency and quality checks rather than vendor leaderboard claims alone.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment