Underdog has released Woof, an English text-to-speech model whose two-voice weights file on Hugging Face is 40.6 MB, about 8.1 times smaller than the 327.2 MB weights file of the open Kokoro-82M model, by our arithmetic from the two repositories’ file sizes. TechCrunch reported the release on 6 November, alongside a browser game built on the model.
The model card says the one-voice version that runs in the browser is a 31.7 MB download. It also says Woof is “free for non-commercial use, with credit” and that “Commercial use needs a license”.
Business Pill · Distribution Advantage
A short explainer on distribution advantage. It teaches the general idea only and does not discuss this story.
The key insight: As we read the model card and file listings, Underdog has published a voice model whose two-voice weights file is 40.6 MB, about 8.1 times smaller than Kokoro-82M’s, by our arithmetic, while keeping commercial use behind a paid license.
The card describes Woof as “a small English text-to-speech model with two voices: Juno and Mira” that “runs on a laptop CPU with nothing but numpy, from a single 40.6 MB file”. It lists 38.24 million parameters in total: 22.95 million in the speech model and 15.3 million in an audio decoder “adapted from Kyutai’s Mimi codec”. TechCrunch reported 31 million parameters.
Underdog says Woof reaches a “1.42% word error rate” on the Gradium TTS Hard Cases set with its Mira voice, and runs “13.8x faster than real time on one CPU thread” with its Juno voice. The card says every model in its comparison was “run and scored by us the same way”, so these are Underdog’s own measurements.
Download sizes of on-device text-to-speech models under 400 MB, in megabytes, as listed in Underdog’s Woof model card on Hugging Face (November 2026). The Woof two-voice file (40.6 MB) and Kokoro-82M’s weights file (327.2 MB) were checked against the Hugging Face file listings.
How the Sizes Compare
In Underdog’s comparison table, Fish Audio Drama 3 is a 41.5 GB download with a 1.89% word error rate, and Kokoro 82M is listed at 327 MB with 5.17%. The card says “Kitten Nano 0.8 is the only smaller download”, at 27.6 MB, with a 9.81% word error rate.
We checked two of the sizes directly. Woof’s weights file, woof_int8.npz, is 40,599,600 bytes, and Kokoro-82M’s kokoro-v1_0.pth is 327,212,226 bytes, according to the Hugging Face file listings. That makes Woof’s file 12.4% of Kokoro’s, by our arithmetic.
The Game and the License
Underdog posted on X on 5 November: “We made a game for our new frontier text-to-speech model, Woof. 31.7 MB, so small that it runs in your browser”. TechCrunch describes a retro room with a typewriter that speaks whatever users type.
TechCrunch reported that the model is “open for everyone to download”. The license sets conditions: commercial use, including “a company’s work”, needs a license from Underdog, and users may not use Woof or its audio “to train, fine-tune or distill another model that makes speech or audio”.
Underdog’s Other Releases
Woof follows Bark, a 167 MB speech-to-text model, TechCrunch reports. In October, Underdog released Saluki 27B, which it described on X as “almost 7x smaller than its full-precision counterpart” and available “under Apache 2.0”. Hugging Face lists Saluki with 762,716 downloads as of 6 November.
Founder Sigil Wen told TechCrunch: “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”.
As we read the card, the accuracy claims are strong but self-reported: Underdog ran and scored every model in its table itself, on one English test set.
As we read the license, “open” here means downloadable weights with conditions: personal, research and non-profit use is free, while a company’s work needs a license and training other speech models on Woof is barred.
As we read Underdog’s releases, it ships models at very different sizes, from the 31.7 MB browser version of Woof to Saluki 27B, which it released under Apache 2.0.
The analysis says: “Every technology stack has layers. Some are scarce, some are abundant.”
As we read it, Underdog publishes Woof’s weights while licensing commercial use, so the weights are open and the commercial right stays with the company. The analysis does not discuss Underdog.
What Is Not Established
The word error rates and speeds are Underdog’s own measurements and have not been independently reproduced. The card lists 38.24 million parameters while TechCrunch reported 31 million. Woof is English-only, according to the card. We did not contact Underdog.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Underdog’s two-voice Woof weights file is 40.6 MB, about 8.1 times smaller than Kokoro-82M’s 327.2 MB file, by our arithmetic from Hugging Face listings, and the browser version is 31.7 MB. Underdog reports a 1.42% word error rate on its own test, and its license requires a paid license for commercial use.
A note on sourcing. Model details, benchmark figures and license terms are from Underdog’s Woof model card and license on Hugging Face; the benchmarks are Underdog’s own measurements. File sizes for Woof and Kokoro-82M were read from the Hugging Face file listings on 6 November 2026. Background and Sigil Wen’s quote are from TechCrunch’s report of 6 November. The size ratio is our arithmetic. We did not contact Underdog.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
