Conway Research Released Underdog AI for Apple Silicon

The local-first assistant utilizes the Husky inference engine to achieve significant throughput gains on M5 Max hardware.

Updated on Oct. 2, 2026 in Artificial Intelligence

Isometric editorial illustration of a silicon microchip fragment with geometric filaments, representing local AI computation on hardware. — Conway Research released Underdog, a new AI assistant that runs locally on Apple silicon to improve data privacy and processing speed. AI Illustration. Upload story photo >

Live Poll

Do you prefer using AI tools that process data locally on your device instead of the cloud?

Conway Research has released Underdog, a new AI assistant powered by a 4-billion-parameter model called Woof that operates locally on Apple silicon devices. The software runs entirely on the user's device, ensuring data privacy by eliminating the need for a Wi-Fi connection.

Why it matters

By moving inference directly onto the device, this software circumvents the latency and privacy trade-offs inherent in cloud-based AI services. This release demonstrates the growing capability of edge-computing frameworks to handle complex models on consumer hardware.

The system achieves a peak performance of 730 tokens per second on an Apple M5 Max, representing a 4.5-fold increase in throughput compared to the Apple MLX framework. The model architecture relies on a 4-billion-parameter configuration known as Woof.

The players

Conway Research

An AI development lab focused on edge-computing software and inference optimization for local hardware.

The details

Underdog processes tasks through the Husky inference engine, a software layer optimized for local execution on Apple silicon. By avoiding remote servers, the assistant processes all data on-device using Flash-enabled function-edit tasks, which reorganize data streams to minimize latency. This allows for rapid model activation without the need for external data transmission.

October 2, 2026: Conway Research published the benchmark results for the Underdog AI assistant.

The Tech Race

This development pushes the boundaries of edge-based AI by significantly outperforming the Apple MLX framework. It establishes a new benchmark for how effectively small-scale, 4-billion-parameter models can be optimized for local Apple silicon execution.

Users with Mac or iPhone hardware can now run AI tasks without relying on cloud servers or Wi-Fi connections. This enables privacy-focused workflows where data never leaves the local device storage.

The takeaway

This software signals a shift toward high-performance local AI that prioritizes user privacy over cloud dependency. Readers should monitor future benchmark releases to see if these gains in token throughput are maintained across older Apple silicon generations.