MacPaw Merges Liquid AI Models With Local macOS Memory Stack

MacPaw and Liquid AI have partnered to co-develop a specialized technology stack for on-device AI on macOS. This collaboration bypasses cloud latency by integrating compute-optimized foundation models directly with local system memory and inference engines on Apple hardware.

The Local AI Stack

The architecture relies on Liquid AI’s foundation models alongside two proprietary MacPaw components. Elix serves as the dedicated on-device inference engine tuned for macOS. Mnemos acts as the memory and context layer, maintaining persistent state across the file system, calendars, and application data without moving information to external servers.

Component	Role in Stack
Liquid Foundation Models (LFMs)	Compute-optimized reasoning and native task execution
Elix	Dedicated macOS on-device inference engine
Mnemos	Persistent context and local memory layer

If you build privacy-sensitive applications, this separation of inference and memory allows the system to retain deep context without requiring continuous cloud synchronization. The integration follows the recent release of Liquid AI’s LFM2.5-2.6B, a 2.69 billion parameter model built for edge devices. Running through this stack, the model achieves up to 220 tokens per second on Apple’s M5 Max chips, shifting the viability of how to run LLMs locally for complex reasoning tasks.

Native Task Execution

MacPaw’s AI assistant, Eney, will serve as the initial implementation of the complete local stack when it ships later in 2026. Eney previously relied on a hybrid approach for file operations and text transformation. The Liquid AI integration pushes the assistant toward complex reasoning and native task execution directly on the machine.

By operating entirely on Apple Silicon’s Neural Engine, the system provides a robust alternative to native OS features. It offers a broader contextual layer that maintains privacy while processing user data locally. This implementation showcases a practical architecture for how AI agents work when constrained to device-level resources.

Third-Party Integration and Setapp Pricing

MacPaw will open this unified inference stack to third-party developers via its Setapp subscription marketplace. External applications will be able to expose their internal context to Eney. This visibility allows the assistant to read app-specific states and execute cross-app workflows locally.

To support this compute model, MacPaw is transitioning Setapp to a credit-based pricing system for AI operations. This replaces the traditional flat-fee subscription, accounting for the variable complexity and processing overhead associated with persistent local foundation model execution.

For macOS developers, the availability of a unified local AI stack lowers the barrier to building complex automation tools. Evaluate your application’s current cloud dependencies and consider how exposing internal state to a persistent local memory layer could enable autonomous cross-app actions for your users.