MacPaw taps Liquid AI to offer on-device inference to devs building for its app store

MacPaw is moving from a cloud-centric AI assistant to a version that runs entirely on the user’s hardware. To achieve this, it has hired Liquid AI to design a model architecture that matches the target device’s capabilities and to build the Elix inference engine along with a local memory system. For engineers, this means swapping out remote API calls for a library that performs inference locally, which can reduce latency and data exposure. The shift also requires integrating the new SDK and possibly retraining models to fit the hardware-specific architecture.

The partnership promises a developer-facing platform that abstracts both on-device and cloud model access. By exposing a single API that can route requests to local inference or to external providers like Google, developers can simplify their codebases and avoid managing multiple vendor integrations. However, adopting this stack will likely involve licensing fees or revenue sharing with MacPaw and Liquid AI, as well as compliance with any credit-based usage accounting they introduce. Engineers must evaluate the cost of these fees against the benefits of reduced bandwidth and improved user privacy.

Apple already supplies its own on-device models, but Liquid AI claims its models are tuned for higher performance across a range of capabilities. This distinction matters for developers who need more specialized or faster models than Apple’s defaults provide. The new stack may not be compatible with devices that lack the necessary compute resources or operating system support, limiting its applicability to newer hardware. Teams should profile target devices to ensure the Elix engine can run within acceptable power and memory budgets.

SetApp’s experimental credit system ties AI operation cost to task complexity, meaning that heavier inference jobs will consume more credits. For developers, this introduces a variable cost model that must be reflected in pricing or subscription strategies for end users. If a user exhausts their credits, the on-device assistant may fall back to a reduced-functionality mode or require a cloud fallback, which could affect the user experience. Engineers need to implement graceful degradation paths and monitor credit consumption to maintain service continuity.