Conway Research has launched Underdog, a personal AI assistant that runs locally, and said it has received investment from a16z, Khosla Ventures and other firms. a16z has confirmed that it led Conway Research’s seed financing round.
Underdog currently runs directly on Mac and connects to a user’s email and calendar. It is designed to find emails, organize schedules and handle routine tasks. The model runs on-device, so personal data does not have to be handed over to a cloud AI service. It can also keep working without an internet connection.
Why it can run on a regular Mac
According to the company, one reason Underdog can run on a standard Mac is that the model is already small enough. The public Woof model currently has 4 billion parameters, and its 4-bit weights are about 2.37 GB. The team also developed a Husky inference engine tailored to Woof and optimized for Apple Silicon.
In company tests on an M5 Max, the same model outperformed MLX across 16 tasks, with the biggest speed gain reaching 4.5x.
Product direction
Founder Sigil Wen said daily work such as searching email, organizing files and managing schedules does not require calling the strongest model in a data center every time. If a smaller model is good enough and fast enough, a user’s own computer can complete a large share of those tasks.
Underdog is also planned for iPhone, Windows, Linux, Android and NVIDIA devices. Wen described that path as 「Underdog’s Law」, saying frontier AI that needs a data center today could have a chance to move onto personal devices in about half a year.
The team is also studying cases in which an agent could shop directly for users. Its goal is to turn local AI into a personal agent that keeps long-term personal context and carries out tasks on a user’s behalf.