> ## Content Index
> Fetch the complete content index at: https://bytevyte.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Underdog AI Assistant Bets That On-Device Privacy Can Pay for Itself
- URL: https://bytevyte.com/underdog-ai-assistant-bets-that-on-device-privacy-can-pay-for-itself/
- Published: 2026-10-09T18:46:17.000Z
- Updated: 2026-10-09T18:46:17.000Z
- Description: Conway Research's Underdog AI assistant runs a 27B model fully on-device, free and ad-free, funded by a commission on purchases made through it.
- Author: Bytevyte Editorial
- Tags: quick-beats

**Underdog**, the on-device AI assistant built by Conway Research, opened an invite-only beta this week for Macs and Windows PCs. The Underdog AI assistant runs its model entirely on hardware the user already owns, keeps personal data away from company servers, and arrives free and ad-free, with Conway planning to charge a small commission on purchases made through the assistant rather than a subscription.

Founder Sigil Wen, a Thiel Fellow who previously worked alongside Perplexity's Aravind Srinivas and OpenAI's Noam Brown, has pitched the product as a direct alternative to cloud-reliant assistants such as Instinct and Muse. Conway Research states that it runs no servers holding personal data and that account keys for email and other connected services are encrypted on the device.

## What the Underdog AI Assistant Does

The assistant covers routine work: email triage, scheduling, meeting notes, dictation, private iMessage, and online errands. Each action sits behind user approval. Underneath sits Husky, an inference engine Wen built to run a 27-billion-parameter reasoning model on consumer hardware by limiting how much data moves between the main processor and the graphics chip.

Conway Research claims the Underdog 27B model outperforms Claude Opus 4.6 Max on most benchmarks while still running on a MacBook or an iPhone. Availability is narrow at launch: Apple-silicon Macs and Windows 11 PCs today, with Linux, iPhone, iPad, Android, Omarchy, and NVIDIA device builds on a waitlist. The company's first funding round is led by a16z, with Khosla Ventures, Hummingbird VC, and Anthology participating.

| Attribute         | Detail                                                   |
| ----------------- | -------------------------------------------------------- |
| Launch status     | Invite-only beta, opened this week                       |
| Platforms today   | Apple-silicon Macs, Windows 11 PCs                       |
| Platforms planned | Linux, iPhone, iPad, Android, Omarchy, NVIDIA devices    |
| Price at launch   | Free, no ads                                             |
| Revenue model     | Small commission on purchases made through the assistant |
| Inference engine  | Husky, built by Conway Research                          |
| Model             | 27-billion-parameter reasoning model                     |
| Backers           | a16z (lead), Khosla Ventures, Hummingbird VC, Anthology  |

## How the Privacy Claim Is Enforced

The technical claim rests on three parts. Conway Research operates no servers that hold personal data, so there is no central store to breach or mine. Keys for email and other accounts a user connects are encrypted on the device, which lets the assistant act on an inbox without the credentials leaving the machine. Every action, from sending mail to running an errand, is gated behind user approval instead of firing autonomously in the background.

Husky exists to make that practical. A 27-billion-parameter reasoning model would normally be too heavy for a laptop, and the engine's design cuts the traffic between the main processor and the graphics chip, the bottleneck that usually decides whether a model runs locally or gets shipped to a data centre. Conway Research also lists NVIDIA devices among its planned targets, alongside the Linux, Android, iPad, and Omarchy builds sitting on the waitlist.

## The Business Model Is the Actual Test

Free is easy to announce. The harder question is what replaces the money cloud assistants collect. A cloud assistant pays for inference on every request, which pushes it toward a subscription, advertising, or both. Running the model locally removes that per-request bill, and that single accounting change is what lets Conway Research offer the Underdog AI assistant at no charge without an ad load.

The replacement is a transaction commission. When a user buys something through the assistant, Conway takes a cut, processed through Stripe. That shifts which user behavior the company gets paid for. Advertising rewards attention and data collection; a purchase commission rewards completed transactions. Neither is neutral, and the commission gives Conway a standing reason to steer users toward buying.

The size of the cut is not public. That figure decides whether the model works. At a small take rate, Conway needs either heavy purchase volume per user or a cost base close to zero, and on-device inference is what makes the second half of that equation possible. A user who never shops through the assistant generates almost no revenue for the company, which is real exposure for a free product.

The commission has a precedent in payment economics, where a small percentage on each transaction funds the rails that carry it. Conway Research applies that logic to an assistant, so its revenue scales with commerce rather than with time spent or data harvested. That alignment is the strongest argument the company has, and it is also the main risk: an assistant that earns nothing from a user who never buys through it has to keep that user cheap enough to carry.

## Where the Trade-Offs Land

For consumers, the arithmetic is straightforward. No subscription, no ads, and no personal data leaving the machine is a better deal than most cloud assistants offer, provided the local model is good enough for daily work. The costs are hardware and platform limits. Windows 11 PCs vary widely in memory and graphics capability, and a 27-billion-parameter model needs headroom that not every machine has. Mac users on Apple silicon get the cleanest path at launch.

Model quality is the second constraint. Conway Research's benchmark claim places Underdog 27B above Claude Opus 4.6 Max on most tests, and independent verification of that claim is not yet available. If the claim holds on real machines, the gap between local and cloud assistants narrows sharply. If it holds only on selected benchmarks, users will notice on longer, messier tasks.

Timing helps the pitch. Consumer hardware in 2026 carries models that needed a data centre two years ago, and the wider market is moving the same direction. NVIDIA's $4,999 DGX Spark desktop and Ethereum's zkAPI work on private inference payments both point to demand for AI that runs outside a hyperscaler's building, and Conway Research is betting the assistant layer follows the same path.

Competitors face a different pressure. Instinct, Muse, and the major platform assistants carry cloud inference costs that the Underdog AI assistant does not, and they cannot match a free, ad-free price without giving up their revenue. What they still hold is distribution: an assistant already installed on a phone or a laptop is harder to displace than one sitting behind a waitlist.

## Why This Matters

Underdog is a live experiment in whether privacy can be a business model rather than a premium feature. Every mainstream assistant has accepted some form of the trade: data, attention, or a monthly fee in exchange for intelligence that runs somewhere else. Conway Research is testing whether local inference strips out enough cost to break that trade, and whether a purchase commission can replace it without recreating the incentive problems it claims to avoid. The commission rate, the pace of the iOS and Android builds, and whether the local model holds up on ordinary Windows hardware are the three things that will settle the question.

## Related Articles

- [Portable Computer: Perplexity and NVIDIA take AI local](https://bytevyte.com/portable-computer-perplexity-and-nvidia-take-ai-local/)
- [Perplexity Hybrid Compute takes the sensitive parts of Mac agent tasks offline](https://bytevyte.com/perplexity-hybrid-compute-takes-the-sensitive-parts-of-mac-agent-tasks-offline/)
- [The Thomson Reuters Thomson model: a $40M bet on owning AI instead of renting](https://bytevyte.com/the-thomson-reuters-thomson-model-a-40m-bet-on-owning-ai-instead-of-renting/)

✔Human Verified

---

*Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*