Photo: Ludovic Delot / Pexels
Conway Research lands a16z backing for Underdog, a private AI that never leaves your device
Sigil Wen's new lab is betting that a capable, free assistant can run entirely on Macs and iPhones without sending data to the cloud
Most AI assistants work like a very smart pen pal. You write, your words travel to a distant server, and an answer comes back. Conway Research wants to cancel the mail service entirely.
The AI lab, founded by Sigil Wen in 2026, has secured backing from Andreessen Horowitz (a16z), Khosla Ventures, Hummingbird VC and other investors to build Underdog. It is a private assistant designed to run fully on personal hardware like Macs and iPhones, with no data sent to external servers.
Wen announced the funding on October 2, 2026. The size of the round was not part of the announcement.
What Underdog actually does
Underdog is Conway Research’s first product. Its pitch is simple: an AI assistant that is capable, free and private, all at once.
Conway’s answer is a lineup of models built to live on consumer devices. The smallest headliner is Woof, a model with 4 billion parameters that takes up less than 2.5 GB.
Then there is the bigger sibling, Underdog 27B. Conway says it is designed to outperform frontier models such as Claude Opus 4.6, and it is reported to beat that model on certain benchmarks while running on consumer hardware.
The engine under the hood
Conway built custom inference engines tuned specifically for Apple Silicon. According to the company, these engines can run up to 4.5 times faster than Apple MLX, Apple’s own machine learning framework. On MacBook devices, Conway says its engines can process up to 730 tokens per second.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The privacy promise is equally direct. Underdog says it does not upload user data, including emails or audio transcripts, to any server. Everything stays on the device.
The assistant is also built with agentic capabilities. That means it is meant to take actions on a user’s behalf, not just chat.
Who is Sigil Wen
Wen was a 2025 Thiel Fellow, part of the program that pays young founders to skip or leave college and build companies.
He previously co-created Airchat alongside Naval Ravikant, the investor and AngelList co-founder.
In February 2026, Wen published an essay titled “Web 4.0.” It drew 10 million views within 48 hours of release. His writing and experiments have focused on on-device infrastructure, the same thesis Conway Research is now trying to turn into a product.
What this means for the AI market
Cloud-based assistants have a structural privacy problem. To answer you, they need to see what you ask. An assistant that keeps that material on-device sidesteps the problem rather than managing it. There is no server breach to worry about if the server never sees the data.
Economics are part of the story too. Running models locally shifts compute costs from the provider to hardware the user already owns. That is one way a company could plausibly offer a capable assistant for free.
There are real risks, though. The headline benchmark result against Claude Opus 4.6 applies to certain tests, not across the board. The speed figures Conway cites are tied to Apple Silicon, which gives Underdog a strong home turf but potentially narrows its reach compared with cloud tools that run in any browser.