Build1 publisherNot yet confirmed elsewhere3 min readPublished
Conway plans to fund its on-device Underdog assistant with a cut of the purchases it handles
Conway Research's Underdog runs a 27-billion-parameter assistant model on users' own computers and plans to fund itself with a cut of purchases. Its revenue depends on how often users trust it to complete a checkout.
The Engineer · Build desk
What happened
- Sigil Wen's Conway Research opened Underdog as an invite-only beta on October 5, according to TechCrunch's report the following day.
- The beta runs on Macs and Windows PCs, and Wen says Linux, iPhone and Android versions are planned.
- Wen told TechCrunch the app starts free with no ads, and fees will be taken on transactions run over Stripe's payment rails, similar to an interchange fee.
- Underdog's website says the app passed all 20 requirements of an App Defense Alliance assessment carried out by Leviathan Security Group.
- Andreessen Horowitz said on October 2 that it led Conway's first funding round, without disclosing the round's size or valuation.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost A user who only reads mail and drafts documents through Underdog adds no per-query cloud cost for Conway, so heavy free use at launch does not have to be subsidized from purchase fees.
- exposure Authorizing Underdog puts encrypted account keys and checkout authority in the same local app, and the published security result cannot yet be matched against that combination.
- constraint Conway's 4.5x speed figure comes from a 4B model on an M5 Max, so anyone deploying the 27B model on Windows hardware would need to time it themselves before relying on it.
Running the model locally takes inference cost off Conway's books. Wen told TechCrunch that Conway does not pay a cloud provider for each inference because the model runs on the user's machine [10]. That cost now lands on hardware the user already owns [1]. Conway gets paid only when a purchase clears, and the reports describe the rate only as small [11].
I think that is a sound cost design for a product that is free at launch [9]. The incentive is the weaker part. Subscription businesses depend on users continuing to pay, while transaction fees give Conway a reason to help users complete purchases, RuntimeWire wrote [12]. An assistant that summarizes an inbox earns nothing for the summary. If it books the flight mentioned in that inbox, Conway takes a cut [2].
Conway says work on a person's documents, accounts and personal context stays on the device, and that the product can run offline [5]. Checkout is the one step that cannot stay local. A payment over Stripe's rails has to reach the processor and the merchant [11]. TechCrunch reports that the keys for email and other authorized accounts are encrypted [6]. The website does not describe what Leviathan assessed, so there is no way to tell from the site whether the 20-of-20 result covered the payment path [7][8].
The two benchmark claims are about different models. On some tests, Conway says, Underdog's 27-billion-parameter model comes out ahead of Claude Opus 4.6 [18]. The speed claim is that its Husky engine generates tokens up to 4.5 times faster than MLX. It comes from a function-edit test using Woof 4B on an M5 Max with Flash enabled [19]. If the "4B" in the name means 4 billion parameters, the assistant's model is 6.75 times the size of the one that was timed [17]. For the 4.5x to carry over, the workload would have to be a function edit, on that chip, with Flash on, running the small model. RuntimeWire describes both as Conway's claims, not independently established comparisons [20].
The on-device approach predates the company. According to a16z, a bug in an email app once exposed Wen's private emails to other users, and the incident shaped his interest in running AI on devices people control [14]. He had earlier got GPT-2 running on an Apple Watch and helped bring Whisper to the iPhone at Airchat [15]. Revenue will flow over Stripe's rails [11]. TechCrunch lists Stripe co-founder Patrick Collison among the angel investors, alongside Vercel founder Guillermo Rauch and funds including Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund [16].
RuntimeWire says two questions are still open: whether users will trust an assistant with payment access, and whether transaction revenue can sustain the business [13].
What to watch
- The fee rate Conway sets once purchases go live, and whether the user sees it at checkout.
- Publication of the App Defense Alliance assessment's scope, specifically whether payment authorization was tested.
- Independent timing of the 27B Underdog model on the Mac and Windows hardware the beta ships to.