Thirty-point-seven million parameters, a WebGPU path in a browser tab, 18x real-time speed on one CPU thread. That is the full spec Underdog prints for Woof, its text-to-speech model.

The architecture is a recombination, not a new design. The audio decoder follows Kyutai's Mimi, and the training set is LibriTTS-R, credited to Yuma Koizumi and colleagues.

The license is where the model stops being useful to most buyers. CC BY-NC 4.0 makes the weights readable but not shippable in a commercial product, and Underdog has not published a separate agreement.

No GPU-path latency, no MOS score, no cloud-TTS comparison, and no hardware floor beyond "single laptop CPU thread" and "most devices."

The draft repeated the license constraint and the architecture note in separate paragraphs, which made the piece feel like a checklist read twice. The rewrite leads with the size-and-license pairing as the single governing fact, folds the architecture into one tight paragraph, and ends on the absence list as a one-sentence closer. The lede reframes the story around what the license permits rather than what the model does, which is the actual news for a developer audience.