Sigil Wen’s Underdog AI Promises Privacy Through Local Inference

The architecture hinges on Husky, a proprietary inference engine that minimizes data transfer between a computer’s main processor and its graphics chip. While Underdog utilizes a 27-billion parameter model derived from Qwen3.8, Wen maintains that it matches the performance of top-tier assistants from six months ago, such as Claude Opus 4.6. This allows the software to manage standard tasks like research or complex problem-solving without transmitting sensitive information to external servers.

Unlike competitors that rely on subscriptions or data monetization, Conway Research plans to sustain Underdog through a fintech-inspired revenue model. By leveraging Stripe’s payment rails, the company intends to collect a small percentage of transaction fees initiated by the AI. This approach aligns the assistant’s financial success with the user’s utility rather than their personal data. With backing from Andreessen Horowitz, Khosla Ventures, and angel investors including Stripe’s Patrick Collison, Wen is positioning the project as a defensive layer against the privacy trade-offs common in the current AI gold rush.