Underdog Launches Private On-Device AI With Stripe Fee Model
Sigil Wen's Underdog runs a 27-billion-parameter AI model on personal devices, stays ad-free, and plans to earn tiny fees only from Stripe transactions.
Summary
Thiel Fellow Sigil Wen’s Conway Research launched Underdog’s invite-only beta on Monday, October 5, 2026. The free, never-ad-supported assistant runs wholly on Macs and Windows PCs, keeping data local, with Linux, iPhone and Android versions coming. Wen says its Husky inference engine moves less data between main and graphics chips than rival local engines, while credentials for authorized email and other accounts are encrypted. Underdog uses a 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B. Wen says some benchmarks compare favorably with Claude Opus 4.6, considered top performance six months earlier, making shopping research and math homework practical.
With users supplying computing power, Conway avoids hosted-inference costs and initially needs neither subscriptions nor ads. It plans to collect a tiny percentage of assistant-initiated payments through Stripe’s secure rails, similar to interchange. The model is intended to avoid mining intimate medical, financial and children’s data, unlike policies permitting collection for advertising, third-party sales or model training. Stripe co-founder Patrick Collison is an angel investor. Other backers include Andreessen Horowitz through Chris Dixon, Khosla Ventures, Hummingbird, SV Angel, the Anthology Fund from Menlo Ventures and Anthropic, Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown and Deedy Das.
Self-taught Wen moved to Silicon Valley at 17 and lived with Andrej Karpathy in an AI hacker house alongside future Perplexity founder Aravind Srinivas and Brown. He tested precursors to Anthropic co-founder Ben Mann’s Claude, David Holz’s Midjourney, OpenAI’s GPT-3 and Stable Diffusion, ran GPT-2 on an Apple Watch, and worked on Naval Ravikant’s now-defunct Clubhouse rival Airchat. He later became a Peter Thiel fellow, choosing projects over college.
Positives
- A 27-billion-parameter Qwen3.8-27B derivative runs entirely on users’ hardware, keeping assistant data off hosted inference systems.
- Mac and Windows support is available in beta, with Linux, iPhone and Android versions planned.
- Free initial access, no advertising and user-supplied computing eliminate immediate subscription pressure and hosted-inference costs.
- Encrypted account credentials add protection when Underdog accesses authorized email and other services.
- Stripe transaction fees could fund the assistant without selling user data or using it to train other models.
Risks & concerns
- Underdog uses a much smaller model than state-of-the-art systems hosted in data centers.
- Capability comparisons rely on Wen’s claim that some benchmarks match Claude Opus 4.6, with no independent results provided.
- Invite-only beta access limits availability, while Linux, iPhone and Android launch dates remain unspecified.
- Useful assistance may require access to sensitive medical, financial and children’s data, raising substantial privacy stakes despite local processing.
- The transaction-fee business model remains unproven and depends on users allowing Underdog to make purchases through Stripe.