Key takeaways
- Underdog released Woof, a 32MB text-to-speech model running in web browsers without internet, containing 31 million parameters.
- Founder Sigil Wen emphasized the company's focus on privacy and local data ownership as alternatives to cloud-dependent AI services.
- Underdog's product line spans multiple scales, from compact Woof to a 27-billion-parameter model for laptops and a waitlisted application.
- The demo game invites users to type text heard back through Woof, while its fictional narrative critiques cloud-dependent AI systems.
Underdog, a startup building AI models designed to run on personal devices, has released an interactive game showcasing its latest text-to-speech model, Woof. At just 32 megabytes in size, the model operates entirely within a web browser without requiring any internet connection, demonstrating the feasibility of deploying speech synthesis locally on hardware most users already own. The demo marks an effort to prove that sophisticated AI capabilities can function without reliance on cloud infrastructure or external services.
Inside the Woof Demo Game
Gameplay and Interaction
The demo takes the form of a game set in a single room decorated with retro-technology aesthetics. The environment includes a typewriter, a cathode-ray tube monitor, and a desk lamp—each interactive element. Players type text into the typewriter using their keyboard, and Woof immediately synthesizes audio, reading the text aloud. The game includes an additional interactive feature: a tuner that can generate different varieties of music when adjusted.
Design Philosophy and Messaging
The game’s introduction contains a warning to players about “intrusive” cloud providers in the fictional town where the game is set. This framing serves as a design statement against cloud-dependent AI systems. The demo’s objective requires players to defeat an antagonist called Darkcloud, and victory unlocks an invitation to try additional Underdog products. All audio synthesis throughout the game relies on the Woof model, demonstrating its consistency across different text inputs and use cases.
Technical Specifications and Performance
The Woof model operates with 31 million parameters and occupies 31.7 megabytes of storage space. According to Underdog, this size allows it to run on most contemporary devices without internet access. The company has released the model openly so anyone can download and deploy it locally on their own hardware. While such a small model inherently cannot match the capabilities of larger, more heavily-trained systems, Underdog identifies two specific applications where Woof’s capabilities prove sufficient: powering local assistant experiences and generating personalized podcasts.
The Emerging Edge AI Landscape
The announcement reflects a growing divergence in AI development approaches. While most well-known AI initiatives have pursued larger models requiring substantial computational resources, a segment of researchers and companies have moved in the opposite direction, training smaller models that can operate on personal devices even when no internet connection exists. This approach prioritizes data privacy, since inference happens locally rather than on servers, and reduces dependence on cloud infrastructure. For users, this means AI capabilities embedded in their phones, laptops, or tablets without sending data externally or incurring latency from network requests.
Underdog’s Broader Product Vision
Previous Releases and Current Focus
Woof is not Underdog’s first locally-deployed model. The company previously released Bark, a speech-to-text model weighing 167 megabytes and designed for transcription and dictation tasks. Underdog has also developed a 27-billion-parameter model aimed at laptop deployment, marking a different point on the scale spectrum. Beyond these specific models, the company is operating a waitlist for an application it describes as a locally-hosted alternative to existing services like Muse and Instinct.
Motivation and Philosophy
Sigil Wen, Underdog’s founder, explained his reasoning to TechCrunch. Rather than adopting existing notetaking applications like Granola, Wen chose to build his own system using models that execute on users’ devices. This approach lets users retain ownership of their data rather than routing it through external services. Wen stated: “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.” The founder has documented this philosophy in what he describes as his manifesto.
The Wider Context: Scale versus Efficiency
The AI industry has traditionally equated progress with scale. The largest and most celebrated models train on the most data and operate on the most powerful hardware. However, computational resources are not infinite—data centers face power constraints, training costs remain substantial, and privacy regulations increasingly restrict how personal data can be used. Against this backdrop, approaches like Underdog’s represent a response: maintaining useful capabilities while dramatically reducing the hardware and bandwidth requirements. Rather than competing on absolute model size, these companies compete on the ratio of capability to resource consumption.
Underdog’s releases over the past months suggest the company is testing different points along this spectrum. The 31-million-parameter Woof addresses the extreme end of size constraints; the 27-billion-parameter model targets devices with more resources; the waitlisted application aims to provide a complete alternative to commercial services. This portfolio approach allows the company to serve different user needs and device capabilities, all built on the principle that computation should happen locally where possible.
What This Means for Users and Developers
For individual users, these releases point toward a future where AI assistance is available without sending data to cloud services or paying per-request fees to external APIs. For developers building applications, especially those targeting privacy-conscious users or regions with limited or unreliable internet, models like Woof and Bark provide functional alternatives to cloud-dependent systems. The game demo itself serves as both a marketing vehicle and a genuine test of Woof’s capabilities. If the model can handle arbitrary user input, synthesize intelligible speech, and maintain performance within a web browser, it proves the underlying technical approach works in practice.
Source: TechCrunch
Frequently Asked Questions
What is Woof and how large is the model?
Woof is Underdog's text-to-speech model with 31 million parameters weighing 31.7 megabytes, small enough to run in web browsers without internet access.
Why does a 32MB speech model matter?
Most speech synthesis relies on cloud infrastructure, whereas Woof demonstrates speech synthesis can function on personal devices locally, challenging the necessity of internet-dependent systems.
What other models has Underdog released?
The company released Bark, a 167MB speech-to-text model for transcription and dictation, plus a 27-billion-parameter model for laptops. It is also operating a waitlist for an application positioned as a locally-hosted alternative to Muse and Instinct.
