Loading page…
The assistant will initially be free and ad-free, its founder says. Its proposed business model takes a share of purchases rather than charging for cloud inference.
Listen to this story
Underdog’s invite-only beta makes local inference both its privacy pitch and cost-control strategy: a 27-billion-parameter model runs on users’ Macs and Windows PCs, with offline use and encrypted keys for connected accounts. Conway Research founder Sigil Wen plans to earn revenue from purchases made through Stripe rather than ads, while Linux, iPhone, and Android support remains planned, not available. Wen’s comparisons with Claude Opus 4.6 are limited to some benchmarks and do not establish broad parity with leading hosted systems.
Wen built Husky, an inference engine he says moves less data between a computer’s main chip and graphics chip to speed local execution.
Underdog’s reasoning model is fine-tuned from Qwen3.8 27B; its reported benchmark comparisons with Claude Opus 4.6 are not independently verified.
Wen proposes taking a small percentage of purchases made through Stripe, avoiding per-request model-provider costs and advertising.
Mac and Windows users have a new, invite-only option for AI assistance that runs its model on their own computer rather than in a data center. Sigil Wen launched Underdog’s beta on Monday, TechCrunch reported. The Conway Research product pairs local processing with Wen’s promise of initially free access and no ads.
The privacy pitch starts with where the model works: Underdog runs wholly on-device, keeping the data it processes on hardware the user owns. It also encrypts the keys used to access email and other accounts that users authorize it to connect to.
Wen’s motivation is personal, according to investor Andreessen Horowitz. In an October 2 post, a16z said a bug in an email app had exposed his private, privileged emails to other users. That experience helped shape his focus on data protection. The investor describes Underdog as designed to work even offline.
“Why should using AI require surrendering your private information?”
Sigil Wen, in his AI manifesto, quoted by TechCrunch
Underdog currently uses a 27-billion-parameter reasoning model fine-tuned from Qwen3.8 27B. Its model is much smaller than today’s leading systems hosted in data centers.
To run it, Wen built Husky, an inference engine—the software that executes an AI model. He says Husky moves less data between a computer’s main chip and graphics chip than other on-device engines. That is his explanation for how the engine is designed to run quickly.
Wen argues that the model compares favorably with Claude Opus 4.6 on some benchmarks. He says it can handle everyday requests such as shopping research and math homework questions. That limited comparison is not an independently verified equivalence across tasks.
Local processing underpins Wen’s commercial plan. Because the AI runs on users’ machines, he says Underdog avoids paying a provider to run the model for each request.
Instead, he plans to take a small percentage of payment transactions the assistant makes through Stripe’s payment system. It is a proposed revenue model tied to purchases made through the assistant, alongside his commitment that Underdog will never carry advertising.
Linux, iPhone and Android versions are described as coming soon; they are planned extensions of the beta’s reach, not platforms available at launch.
Loading discussion...
Join the conversation
Explain whether that revenue model would affect your confidence in its recommendations.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.