Wen, who moved to Silicon Valley at 17 and lived with figures such as Andrej Karpathy, Aravind Srinivas, and Noam Brown, built Underdog on a custom inference engine called Husky. According to Wen, Husky moves less data between the computer's main chip and its graphics chip, enabling faster on-device performance.
Underdog currently uses a 27-billion parameter reasoning model fine-tuned from Qwen3.8 27B. Wen argues that this model compares favourably in some benchmarks with Claude Opus 4.6, or what was considered top performance six months ago. "You don't need to sacrifice your privacy for the capability because they're just as capable," he said.
The assistant's business model is unusual for the AI industry. Underdog will be free initially and never ad-supported. Because the AI runs on users' hardware, Wen said the service has minimal inference costs: "I don't have to charge you a subscription to run this because my costs are so super low." Instead, backed by angel investors including Stripe co-founder Patrick Collison, Underdog will take a tiny percentage of payment transactions made through the assistant using Stripe's rails, similar to an interchange fee. This approach, Wen says, ensures the assistant does not need to mine user data.
This strategy contrasts with many competing AI assistants, whose privacy policies can allow them to collect data on users. Underdog's security features include encrypting the keys to the email and other accounts users authorise it to access.