A powerful AI assistant does not have to send your data away. Underdog, launched by Sigil Wen on October 6, 2026, runs wholly on-device, bringing a 27-billion parameter reasoning model to devices users already own.
The project arrives with an ambitious promise: combine serious AI capability with privacy, account security, and a business model that does not depend on advertising. Underdog will be free at first, never ad-supported, and designed to keep user data on-device.
A 27B Reasoning Model in a 7.89 GB Package
Underdog currently uses a model fine-tuned from Qwen3.8 27B, giving the assistant 27B parameters and a 27-billion parameter reasoning foundation. The model has a file size of 7.89 GB, putting a 27B-class agent model on hardware users already own.
That combination makes Underdog stand out in a crowded AI assistant market. Instead of relying on a remote service for every interaction, the assistant runs wholly on-device, so the user’s data remains on devices they already own. The local approach sits at the center of Underdog’s identity rather than serving as a side feature.
Wen says the model compares favorably with Claude Opus 4.6 in some benchmarks. That claim gives the project an intriguing edge: Underdog is pairing a small, privacy-focused model with performance that can stand alongside a major model in selected tests.
The result is a clear challenge to the assumption that users must trade privacy for capability. “You don’t need to sacrifice your privacy for the capability because they’re just as capable,” Wen said.
Privacy Built Around Authorized Accounts
Underdog’s privacy design extends beyond keeping model activity on the device. The assistant includes security features that encrypt the keys to email and other accounts users authorize Underdog to access.
That detail matters because an assistant becomes more useful when it can work with accounts, but access also creates a serious security responsibility. Underdog’s approach puts encrypted account keys at the heart of that relationship, while the assistant continues to run wholly on-device.
The model is small enough to fit in a 7.89 GB file, yet it belongs to the 27B-class agent model category. That pairing connects local operation, account access, and reasoning capability in one product. Users are not being offered privacy as a stripped-down version of an online assistant; Underdog is built around privacy from the start.
Wen, the creator of Underdog and an AI hacker, frames the project as a personal mission rather than a short-term experiment. “I honestly want to build Underdog for myself. I’m building a product that I would be proud for my future children to use,” Wen said.
A Free Start With Payments Behind the Assistant
Underdog will be free at first and never ad-supported, giving the assistant a different path from products that turn attention into revenue. The company behind the project, Conway Research, plans to earn a tiny percentage of payment transactions that the AI assistant makes using Stripe’s secure payment rails.
That model links the assistant’s usefulness to the actions it can take for users. Instead of placing ads inside the product, Conway Research will take a tiny percentage from payment transactions made through Stripe’s secure payment rails. The assistant remains free at first, while the business receives revenue when it handles those transactions.
Conway Research has financial backing from Andreessen Horowitz through partner Chris Dixon, along with Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund. Angel investors Guillermo Rauch, Noam Brown, and Deedy Das also back the startup.
That investor group gives Underdog a strong launch platform, but the product’s core test will happen on users’ devices. Can a 27-billion parameter reasoning model deliver useful assistance without collecting data or showing ads? Can encrypted account keys and Stripe’s secure payment rails support an assistant that acts on a user’s behalf?
Underdog begins with a focused formula: a 7.89 GB local model, 27B parameters, privacy-centered operation, and a payment plan that avoids advertising. If the assistant lives up to its benchmark comparisons and Wen’s vision, it could push AI assistants toward a future where capability and privacy move together instead of pulling users in opposite directions.
Based on
