Conway Research debuts Underdog private on-device AI assistant

Conway Research has launched an invite-only beta of Underdog, an AI assistant designed to run wholly on a user’s device rather than in a remote data centre. The application is available initially for Macs and Windows PCs, with Linux, iPhone and Android versions planned. It currently uses a 27-billion-parameter reasoning model fine-tuned from Qwen3.8 27B.

Founder Sigil Wen says the product is intended to offer assistant functions such as shopping research and answers to maths homework questions without requiring users to send their information to a hosted model provider. Wen, a Thiel Fellow, built the startup around the premise that useful assistants may require access to highly sensitive personal information.

Local inference and account protection

Underdog is powered by Husky, an inference engine built by Wen to run AI models quickly on local hardware. He says Husky is designed to move less data between a computer’s main chip and graphics chip than other on-device engines. That architecture is central to the product’s claim that user data remains on devices the customer already owns.

The assistant also encrypts keys for email and other accounts that a user authorises it to access. Those protections matter because an AI assistant can be asked to work with medical details, financial information and data about family members. Underdog’s local approach contrasts with services whose policies can permit the collection of user data for advertising, sharing with third parties or model training.

A smaller model and a different revenue plan

Underdog does not use the largest frontier models hosted in data centres. Wen says the current 27B model compares favourably with Claude Opus 4.6 in some benchmarks, while acknowledging that it is smaller than today’s state-of-the-art hosted systems. His argument is that the model is sufficient for everyday assistant tasks and that local models will improve over time.

The company plans to make the app free initially and says it will not be supported by advertising. Because inference runs on customer devices, Wen says Conway Research avoids the large provider costs associated with serving models from the cloud. The proposed business model is instead to take a small percentage of payment transactions the assistant makes through Stripe’s secure payment rails, similar to an interchange fee.

Backers and business implications

Conway Research is backed by Andreessen Horowitz through partner Chris Dixon, Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund, a partnership fund of Menlo Ventures and Anthropic. Angel investors include Stripe co-founder Patrick Collison, Vercel founder Guillermo Rauch, Noam Brown and Deedy Das.

For businesses evaluating AI assistants, Underdog highlights a practical procurement question: whether sensitive work can stay on employee devices, and how the provider’s revenue model may shape data handling. Teams should compare local-model capability, authorised-account controls and payment workflows against the tasks they intend to delegate before enabling an assistant for sensitive use.