Three-Layer Architecture: LFMs + Elix + Mnemos

The partnership combines Liquid AI's models with two existing MacPaw technologies into a three-layer architecture designed entirely for Apple Silicon :

• Liquid Foundation Models (LFMs) — Liquid AI is designing specialized small models optimized for macOS AI assistance, balancing efficiency under tight compute budgets with high performance . These models are built for "efficiency under tight resource constraints," allowing them to achieve quality comparable to much larger models while running on a laptop .
• Elix — MacPaw's on-device inference engine, built on Apple's MLX framework and tuned by MacPaw Research for speed and privacy. It runs all model execution locally on the user's Mac .
• Mnemos — A local-first persistent memory layer for AI agents. It maintains a living, bi-temporal knowledge base entirely on-device using a single libSQL file, with secrets stored in a CryptoKit vault behind Touch ID . Nothing leaves the Mac unless the user explicitly routes it out .

Privacy and Offline Advantages

The stack is designed to keep all data on the Mac, never leaving the device. According to the announcement, this means user data never touches external servers, which provides stronger privacy guarantees and full offline functionality compared to cloud-dependent AI solutions . MacPaw's CEO described the stack as a "truly private, local AI layer" that can run AI even without an internet connection .

For context, this approach contrasts with the cloud-reliant architecture of many competing AI assistants. Apple's own AI efforts, while increasingly on-device, do not currently offer the combination of localized persistent memory and native macOS task execution that this partnership targets .

Timeline: On-Device Capabilities Coming Later This Year

Results from the partnership are expected in 2026 . Eney already shifted to a local model in 2025 and now includes reasoning, contextual search, skills, and actions; the partnership accelerates the addition of Liquid's models and the full on-device memory and inference stack . The native AI functions are planned to come to Eney and Setapp "later this year" .

At the time of the announcement, Eney was still cloud-only, with the expectation that models can be run on device with the update expected later in 2026 .

Setapp Developer Plans: Credit-Based Pricing and Open Access

After Eney is tested, MacPaw plans to open the shared on-device inference infrastructure to third-party developers building for its Setapp marketplace, which has over 150,000 paying subscribers . Key details include:

• Credit-based pricing — Setapp will experiment with pricing tied to AI task complexity rather than fixed subscription tiers .
• Open stack — Developers will be able to use the same Elix inference engine and Mnemos memory layer in their own Setapp-distributed apps .
• First testing in Eney — The infrastructure will be proven in MacPaw's own assistant before being exposed to the developer ecosystem . Setapp access is described as a future step after Eney testing and broader evaluation, not an immediately available SDK .

What This Means for Mac Users

For everyday Mac users, the partnership promises an AI assistant that works without sending personal data to remote servers, responds quickly because models run locally, and continues functioning even without an internet connection . The combination of on-device intelligence, persistent memory, and native macOS task execution is one that "hasn't existed on the Mac before," according to MacPaw's announcement .

No financial terms were disclosed; the agreement is a co-development partnership, not an acquisition or equity deal .

Three-Layer Architecture: LFMs + Elix + Mnemos

The partnership combines Liquid AI's models with two existing MacPaw technologies into a three-layer architecture designed entirely for Apple Silicon :

• Liquid Foundation Models (LFMs) — Liquid AI is designing specialized small models optimized for macOS AI assistance, balancing efficiency under tight compute budgets with high performance . These models are built for "efficiency under tight resource constraints," allowing them to achieve quality comparable to much larger models while running on a laptop .
• Elix — MacPaw's on-device inference engine, built on Apple's MLX framework and tuned by MacPaw Research for speed and privacy. It runs all model execution locally on the user's Mac .
• Mnemos — A local-first persistent memory layer for AI agents. It maintains a living, bi-temporal knowledge base entirely on-device using a single libSQL file, with secrets stored in a CryptoKit vault behind Touch ID . Nothing leaves the Mac unless the user explicitly routes it out .

Privacy and Offline Advantages

The stack is designed to keep all data on the Mac, never leaving the device. According to the announcement, this means user data never touches external servers, which provides stronger privacy guarantees and full offline functionality compared to cloud-dependent AI solutions . MacPaw's CEO described the stack as a "truly private, local AI layer" that can run AI even without an internet connection .

For context, this approach contrasts with the cloud-reliant architecture of many competing AI assistants. Apple's own AI efforts, while increasingly on-device, do not currently offer the combination of localized persistent memory and native macOS task execution that this partnership targets .

Timeline: On-Device Capabilities Coming Later This Year

Results from the partnership are expected in 2026 . Eney already shifted to a local model in 2025 and now includes reasoning, contextual search, skills, and actions; the partnership accelerates the addition of Liquid's models and the full on-device memory and inference stack . The native AI functions are planned to come to Eney and Setapp "later this year" .

At the time of the announcement, Eney was still cloud-only, with the expectation that models can be run on device with the update expected later in 2026 .

Setapp Developer Plans: Credit-Based Pricing and Open Access

After Eney is tested, MacPaw plans to open the shared on-device inference infrastructure to third-party developers building for its Setapp marketplace, which has over 150,000 paying subscribers . Key details include:

• Credit-based pricing — Setapp will experiment with pricing tied to AI task complexity rather than fixed subscription tiers .
• Open stack — Developers will be able to use the same Elix inference engine and Mnemos memory layer in their own Setapp-distributed apps .
• First testing in Eney — The infrastructure will be proven in MacPaw's own assistant before being exposed to the developer ecosystem . Setapp access is described as a future step after Eney testing and broader evaluation, not an immediately available SDK .

What This Means for Mac Users

For everyday Mac users, the partnership promises an AI assistant that works without sending personal data to remote servers, responds quickly because models run locally, and continues functioning even without an internet connection . The combination of on-device intelligence, persistent memory, and native macOS task execution is one that "hasn't existed on the Mac before," according to MacPaw's announcement .

No financial terms were disclosed; the agreement is a co-development partnership, not an acquisition or equity deal .