Underdog

AI · Consumer Software · Privacy

Launch

Sigil Wen launches Underdog, a private on-device AI assistant

Running inference locally eliminates data-center costs, letting the startup earn from payment transactions instead of subscriptions or ads.

• Conway Research founder Sigil Wen opened an invite-only beta of Underdog, an AI assistant that runs entirely on a user's own Mac or Windows PC rather than in the cloud.
• Wen is a self-taught coder and Thiel Fellow who, at 17, lived in a Silicon Valley hacker house alongside Andrej Karpathy and tested early prototypes of Claude, Midjourney and GPT-3.
• Underdog runs a 27-billion-parameter model fine-tuned from Qwen3.8-27B on a custom inference engine called Husky, which Wen says minimizes data movement between a device's CPU and GPU.
• The app encrypts the keys to any email or accounts users authorize it to access, and Wen claims its 27B model matches Claude Opus 4.6-level performance on some benchmarks.
• Underdog will be free with no subscription or ads; instead it plans to take a small cut of transactions when users complete purchases through Stripe's payment rails.
• Andreessen Horowitz, via partner Chris Dixon, led Conway Research's first funding round, with Khosla Ventures, Hummingbird, SV Angel and the Anthropic-Menlo Anthology fund also participating; terms were undisclosed.
• By shifting inference from data centers to users' own hardware, Underdog removes the server costs that force rivals into subscriptions or ad-supported models, testing whether privacy and capability no longer require tradeoffs.