Underdog launches privacy-focused on-device AI assistant

Sigil Wen, a Thiel Fellow, has launched the private beta for Underdog, an artificial intelligence assistant designed to run entirely on-device. By executing locally on Mac and Windows platforms, the assistant avoids sending user data to external servers, positioning itself as a privacy-centric alternative to existing AI tools.

To optimize performance, Wen developed Husky, a proprietary inference engine designed to minimize data movement between the main chip and the GPU. The system utilizes a 27-billion-parameter reasoning model fine-tuned from Qwen3. 8-27B, which aims to compete with high-end models for everyday tasks such as research and mathematics.

Underdog employs an unconventional business model. Rather than using subscriptions or advertisements, the service will be free to use and will instead collect a small commission on payment transactions processed through secure Stripe integrations. The startup has secured backing from several prominent investors, including Patrick Collison of Stripe, Andreessen Horowitz, and Khosla Ventures.