Key facts
- Underdog is an AI assistant that runs on-device, processing user data locally.
- The app is currently available for Mac and Windows PCs, with Linux, iPhone, and Android versions planned.
- Underdog uses a 27-billion parameter reasoning model fine-tuned from Qwen3.8 27B.
- Wen claims the model performs comparably to Claude Opus 4.6 on everyday tasks.
- Underdog's business model involves taking a small percentage of payment transactions facilitated by the AI.
- Conway Research, the startup behind Underdog, is backed by investors including Andreessen Horowitz and Khosla Ventures.
Sigil Wen, a Thiel Fellow and former resident of an AI hacker house, has launched Underdog, an invite-only AI assistant designed to prioritize user privacy. The application runs entirely on-device, meaning user data remains on their own machines, currently Macs and Windows PCs, with future support for Linux, iPhone, and Android.
Wen developed Husky, an inference engine that powers Underdog, which he claims is designed for speed by minimizing data transfer between a computer's main and graphics chips. Additional security features include encryption of account access keys. Underdog utilizes a 27-billion parameter reasoning model, fine-tuned from Qwen3.8 27B, which Wen asserts is comparable in capability to models like Claude Opus 4.6 for everyday tasks, arguing that privacy does not need to be sacrificed for functionality.
The startup, Conway Research, is backed by notable investors including Andreessen Horowitz, Khosla Ventures, and angel investors such as Stripe co-founder Patrick Collison. Underdog's business model deviates from many AI assistants by not being ad-supported and avoiding large overhead costs associated with cloud-based inference. Instead, it plans to generate revenue by taking a small percentage of payment transactions processed through the AI assistant via Stripe's payment rails, aligning its interests with users similarly to banks or credit card providers.