Underdog's 32MB Woof Speech Model Runs Offline on Most Devices
Underdog has released Woof, a speech generation model that is 32MB in size and that the company says can run on most devices without any internet connection. According to TechCrunch, which reported the release on November 6, 2026, the model has 31 million parameters, the learned numerical weights that determine its behavior, and is open for anyone to download.
The company is demonstrating Woof through a browser-based game at underdog.ai/escape. Users type anything on a virtual typewriter, and the game speaks the text back. A music tuner in the game can be rotated to generate different kinds of music. Because the demo runs in the browser, developers can hear the model's output without installing anything.
Underdog's founder is Sigil Wen. The company's earlier release, Bark, is a 167 MB speech-to-text model built to run on edge devices, meaning local hardware rather than remote servers, for transcription and dictation. Woof works in the opposite direction, turning text into speech, so the two models cover the input and output halves of a voice pipeline. Underdog has also released a 27-billion-parameter model that can run on laptops, which places its work across a wide range of model sizes.
A quick check of the headline numbers is useful for anyone thinking about deployment. At one byte per parameter, 31 million parameters occupy roughly 31MB. A 32MB file therefore implies weights stored at about 8 bits each, with a small overhead for the rest of the file. That points to aggressive quantization, the process of storing weights at lower numerical precision, which is the standard route to fitting a neural network into the memory and storage budgets of phones, microcontrollers, and older hardware. The arithmetic says nothing about output quality, latency on specific chips, or how the model handles languages, accents, or prosody (the rhythm, stress, and intonation of speech). Those are the questions that matter once the weights are in hand.
In my view, the more practical question for engineers is what a model this small makes cheap to test. Offline speech generation has mostly meant either a cloud API, which charges per call and depends on a network, or a large local model that needs a capable GPU. A sub-40MB open-weights option changes that evaluation for game audio, accessibility tools, and embedded voice interfaces, where a network round trip is either unacceptable or impossible.
The broader context also includes the risks that come with the same property. Open weights that run offline cannot be recalled or metered, so misuse, including voice imitation, is harder to limit at the platform level. Those concerns apply to every capable voice model, and a smaller model lowers the hardware bar for everyone, including people with bad intentions. Developers building on Woof will need their own safeguards, such as consent checks and provenance tooling, rather than relying on a hosted service to enforce them.
The release is most useful as a data point on how far speech generation has moved toward the edge. A model that speaks typed text through a browser game, on a device with no connection, is a concrete step in that direction. Whether Woof's quality holds up under real workloads is a question the open download now lets the community answer directly.
