Sigil Wen was 17 when he moved to Silicon Valley and lived in an AI hacker house with the researcher Andrej Karpathy. There he coded alongside people who would become some of the industry's biggest names — Perplexity founder Aravind Srinivas, OpenAI researcher Noam Brown — and tested early versions of tools that later shipped as Claude, Midjourney, GPT-3 and Stable Diffusion. For fun, he got GPT-2 running on an Apple Watch. Now a Thiel Fellow, Wen launched the invite-only beta of Underdog on Monday: an AI assistant that runs entirely on the device.
That puts Underdog against a growing class of local assistants that TechCrunch's headline groups as Instinct and Muse. Underdog currently runs on Macs and Windows PCs, with Linux, iPhone and Android versions promised. User data stays on hardware people already own, the company says.
What is inside
Wen built a custom inference engine called Husky, designed to move less data between the machine's main chip and its graphics chip than other on-device engines — that data movement is usually the bottleneck for local inference. Underdog also encrypts the keys to the email and other accounts a user authorises it to reach.
The model itself is far smaller than the systems running in data centres: a 27-billion-parameter reasoning model fine-tuned from Qwen3.8 27B. Wen argues it compares favourably with Claude Opus 4.6 in some benchmarks — that is, with what counted as top performance six months ago — and that this is enough for everyday tasks such as shopping research or maths homework. "You don't need to sacrifice your privacy for the capability because they're just as capable," he says.
The business model
The more interesting question may be how it plans to make money. The app is free at launch and is not meant to carry ads. Because inference runs on users' machines, there is no provider bill for Conway Research, the company behind Underdog. "I don't have to charge you a subscription to run this because my costs are so super low," Wen said.
Instead he is borrowing a page from fintech: Underdog will take a small percentage of payment transactions the assistant makes through Stripe's payment rails, something like an interchange fee. That way, the argument goes, the assistant never has to mine user data — deliberately the opposite of rivals whose privacy policies allow collecting data that they may sell to advertisers or use to train other models. For an assistant that needs access to medical conditions, financial data or details about a user's children to be useful, that trade-off is unusually consequential.
Who is behind it
Backers include Andreessen Horowitz through partner Chris Dixon, Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund, the joint fund from Menlo Ventures and Anthropic. Angel investors include Stripe co-founder Patrick Collison, Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown and Deedy Das.
What is still open
The beta is invitation-only, and the performance comparisons come from the founder himself; independent benchmarks putting a 27B model on consumer hardware against current cloud models are not yet available. The payment cut is also untested as a revenue model, since it assumes users regularly let their assistant transact. Whatever the outcome, the direction is notable: the question of whether AI assistance must cost privacy is now being answered not only with smaller models, but with a different business model.
