Underdog launches private on-device AI agent with payment-based monetization — no data centers, no subscriptions

Underdog

Sigil Wen is building Underdog, a personal AI agent that runs entirely on-device, with no data centers, no cloud inference, and no subscription fee. The pitch is privacy by architecture: the model comes to your data, not the other way around.

Cross-platform from day one. Wen says Underdog supports Apple products, Windows via Nvidia hardware, Linux, Android, and iOS, with the iOS app imminent. The breadth is intentional — the goal is to reach users on hardware they've already paid for rather than requiring new devices.

“We're building Underdog, which is your private personal AI that runs on the device... Underdog 27b scores better than Claude Opus 4.6 max, which is the frontier model from Anthropic from six months ago... On Friday, we opened to an invite only beta. Over the weekend, more than around 5,000 tweets went out of people asking for invites.”

On-device doesn't mean underpowered. Underdog's core model, a 27B parameter version post-trained with reinforcement learning on top of Qwen's 27B base, scores above Claude Opus 4.6 Max on Artificial Analysis benchmarks. Wen frames this as validation of what he calls "Underdog's Law" — that frontier models become runnable on consumer devices within roughly six months of their release. The team also released an inference engine called Husky, which Wen claims delivers a 445x performance gain, though the baseline for that figure isn't specified in the conversation.

Monetization through payments, not subscriptions. Because there's no server infrastructure, the cost base that forces competitors into monthly fees doesn't exist. Instead, Underdog monetizes through agentic commerce — when the app browses the web for flights, finds the cheapest option, and completes the purchase, it generates a unique card and takes a cut of the transaction. Patrick Collison, CEO of Stripe, is the company's biggest angel investor, which is both a signal of credibility and a practical fit given the payments-based model.

Early traction. Underdog opened an invite-only beta on a Friday, and more than 5,000 tweets requesting invites went out over the weekend. Wen says users are already unsubscribing from WhisperFlow, Granola, and Superhuman, citing on-device privacy as the reason.

The vulnerability of cloud-based AI apps to cyberattacks — Wen cites OpenAI as an example — is the sharpest version of his competitive argument. If on-device models reach parity with cloud counterparts, storing user data in a centralized target becomes harder to justify to consumers. Whether Underdog can sustain that parity as frontier models continue to advance is the open question.