Conway Research Launches On-Device AI Assistant Underdog in Invite-Only Beta
Founded by Thiel Fellow Sigil Wen, the startup runs local models to eliminate server inference costs and safeguard personal data.

Sigil Wen, a Thiel Fellow and veteran of Silicon Valley AI hacker houses, has launched an invite-only beta of Underdog, an artificial intelligence assistant designed to run entirely on local hardware. Developed by Wen's startup, Conway Research, the application keeps user data stored locally rather than transmitting prompts and credentials to centralized cloud data centers, as first reported by TechCrunch Startups (https://techcrunch.com/2026/10/06/silicon-valleys-ai-wunderkind-launches-underdog-the-most-private-instinct-muse-competitor-yet/).
Underdog is currently available for macOS and Windows PCs, with future client support planned for Linux, iPhone, and Android platforms. To run local language models efficiently on desktop machines, Wen built a proprietary inference engine called Husky. According to Wen, Husky accelerates response times by minimizing data transfers between a system's primary processor and its graphics chip. The software also incorporates local encryption for third-party access keys, including connected email accounts.
The desktop application runs a 27-billion-parameter reasoning model fine-tuned from Qwen3.8 27B. While smaller than frontier foundation models running in enterprise data centers, Wen claims the model matches or exceeds the benchmark performance of Claude Opus 4.6 from six months ago across everyday workflows, such as product research and mathematical problem-solving. Wen asserts that small, on-device models will continue to close capability gaps with cloud-hosted alternatives.
Conway Research is pairing its local compute architecture with an unconventional monetization strategy that bypasses recurring consumer subscriptions and advertising. Because client-side execution removes third-party inference compute bills, Underdog plans to remain free to use. With Stripe co-founder Patrick Collison participating as an angel investor, Conway Research plans to generate revenue by taking a small percentage fee—resembling interchange fees—on commercial transactions executed through Stripe's payment rails.
Wen, who previously developed a GPT-2 port for Apple Watch and worked at audio platform Airchat, founded Conway Research after living in an AI-focused hacker house with researcher Andrej Karpathy. The startup has raised capital from Andreessen Horowitz partner Chris Dixon, Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund—a joint vehicle between Menlo Ventures and Anthropic. Individual angel backers include Collison, Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown, and Deedy Das.
Sources
Written by
The Company Wire
Inside the companies building what’s next. Reporting on startups, technology, funding and the people shaping them.


