Sigil Wen launched Underdog on Monday, an invite-only beta for an on-device AI assistant that processes all requests locally on a user's Mac or Windows PC, with plans to support Linux, iPhone and Android.
The pitch is direct: keep data private and avoid the per-query economics of cloud inference. Wen, a Thiel Fellow and self-taught coder, argued that on-device execution solves the privacy problem while eliminating the marginal cost structure that forces cloud-based AI providers to charge subscription fees.
Underdog runs on Husky, an inference engine Wen built to minimize data movement between a computer's CPU and GPU—a key efficiency metric in on-device AI. The assistant uses a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B, which Wen claims matches Claude Opus performance on certain benchmarks, though he acknowledges that bar was set six months ago.
Wen, who lived in an AI hacker house with researcher Andrej Karpathy at age 17 and worked alongside Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown, has positioned himself as early to the technology. He tested prototypes that became Claude, Midjourney, GPT-3 and Stable Diffusion. Naval Ravikant later hired him to code Airchat.
The business model is free, with no advertising planned. Wen said the operational cost of running a local model is low enough that he can absorb it without subscription revenue—a structural advantage over cloud providers whose unit economics depend on either API pricing or paid tiers.
Stripe co-founder Patrick Collison is among Underdog's angel investors, a signal of backing from someone familiar with payment infrastructure and SaaS scaling.
