MacPaw and Liquid AI Build On-Device Inference Stack for Third-Party Developers

A Kyiv-Based Developer Goes Local

MacPaw, the Ukrainian software house behind CleanMyMac and the SetApp subscription platform, is embedding on-device inference into its product line through a technical partnership with Liquid AI. The collaboration centers on Elix, a locally hosted inference engine, and an accompanying memory system that will power MacPaw's own AI assistant, Eney, before rolling out to the wider developer community on SetApp.

At DailyTechWire, we've tracked the steady drumbeat of edge-AI announcements from Cupertino to Shenzhen, but this deal stands out for two reasons: MacPaw intends to open the entire stack to third parties, and it is testing a credit-based metering model that ties pricing directly to computational load rather than flat subscriptions. Both moves suggest the company sees SetApp less as a curated catalog and more as infrastructure for a new class of agentic applications that need to run without a network connection.

Architecture Before Scale

Liquid AI's pitch hinges on tailoring neural-network topology to the silicon it will inhabit. According to Ramin Hasani, co-founder and CEO of Liquid AI, the startup selects architectures optimized for target hardware before any training begins, a workflow intended to squeeze maximum efficiency from the fixed transistor budget of a laptop or phone.

"That allows us to really have the most efficient version of intelligence that runs directly on the device, with benefits like privacy and security," Hasani explained. The claim echoes the design philosophy behind Apple's own Neural Engine and Google's Tensor units, but Liquid AI is positioning its models as performance-tuned for specific capabilities rather than general-purpose building blocks.

The startup is also layering in a customization framework that lets models ingest user data and refine their behavior over time. Hasani frames this as adaptive intelligence, a step beyond static weights frozen at training time. Whether that adaptability translates to measurable accuracy gains in production remains an open engineering question, one that will be answered as MacPaw's Eney assistant moves from preview to general release.

Offline Agentic Workflows Enter the Equation

Oleksandr Kosovan, CEO of MacPaw, emphasized that local execution unlocks offline operation for assistants and multi-step workflows. That matters in markets where connectivity is intermittent or expensive, and it sidesteps the latency tax of round-tripping prompts to a distant data center. It also keeps sensitive data, whether financial records or medical notes, on the user's machine.

Apple already ships its own suite of local models to developers through Core ML and the recently expanded Apple Intelligence APIs. Yet Liquid AI argues there is room for alternatives that prioritize different performance envelopes or offer finer-grained customization hooks. The real test will be adoption: convincing developers to integrate a second inference runtime alongside the platform vendor's own tooling is a high bar, especially when battery life and thermal headroom are shared resources.

SetApp as Inference Marketplace

MacPaw operates SetApp, a subscription service with more than 150,000 paying users who gain access to a rotating library of macOS and iOS applications for a monthly fee. The company now plans to position the platform as a distribution and monetization layer for AI-native apps, complete with backend access to both on-device and cloud models.

Once the Liquid AI integration is stable, MacPaw will expose the inference stack to any developer building for SetApp. Kosovan indicated the platform will also broker connections to remote models from Google and other providers, consolidating endpoint management in a single SDK. That hybrid model, local for latency-sensitive tasks and cloud for heavy lifts, mirrors the architecture patterns emerging across the industry.

The more novel piece is pricing. MacPaw is experimenting with a credit system in which each AI operation consumes a variable number of tokens depending on its computational cost. A lightweight text classification might cost one credit; a multi-turn reasoning chain with memory lookups could cost dozens. Users purchase or earn credits as part of their subscription tier, and developers set their own credit rates within guardrails MacPaw defines.

This approach attempts to solve a problem that has bedeviled SaaS companies grafting generative AI onto fixed-price plans: inference costs vary wildly by workload, and flat fees either leave money on the table or expose the vendor to unbounded compute bills. Credits introduce usage-based economics without forcing end users to parse per-token invoices. Whether consumers will tolerate yet another virtual currency, and whether developers will find the split attractive, will determine if the model spreads beyond SetApp.

The Broader Bet on Edge Intelligence

MacPaw's move is part of a wider industry reorientation toward the edge. Export controls on high-end GPUs have made cloud inference more expensive and geopolitically fraught for companies serving multiple regions. Latency requirements for real-time applications, from voice assistants to code editors, favor local execution. And privacy regulations in Europe and parts of Asia penalize unnecessary data movement.

Liquid AI itself emerged from MIT research into liquid neural networks, a class of continuous-time models originally designed for robotics and control systems. Applying that work to consumer software is a category leap, and the startup will need to demonstrate that its architectural choices deliver tangible wins in accuracy, speed, or energy efficiency. Early partnerships like MacPaw's provide a testbed, but scale will require buy-in from larger platform players or a breakout developer success story.

For MacPaw, the partnership is also a signal of ambition. The company has built a profitable business around Mac utilities, but Eney and the SetApp AI initiative represent a bid to capture a share of the agentic application layer before incumbents lock it down. Whether a 150,000-user subscription base is enough runway to compete with the app stores operated by Apple, Google, and Microsoft is an open question. The answer will hinge on execution: how fast the inference stack ships, how many developers integrate it, and whether credit-based pricing proves to be an elegant solution or another friction point in an already crowded market.

What Comes Next

MacPaw has not disclosed a public timeline for the Liquid AI rollout, but Eney is already in limited preview, and the company has indicated that developer access to the inference stack will follow once the architecture is production-ready. The credit-based pricing model is live in experimental form on SetApp, giving the team real usage data to refine the economics before a broader launch.

The partnership also raises a tactical question for other regional developers: as U.S. and Chinese tech giants dominate the AI stack, can independent software houses in Kyiv, Seoul, or Jakarta carve out defensible positions by assembling best-of-breed components and competing on integration, pricing, or vertical focus? MacPaw's answer appears to be yes, but only if the underlying technology delivers and the go-to-market execution is flawless. The next twelve months will reveal whether on-device inference becomes a commodity or a differentiator, and whether a subscription app store can evolve into a credible AI platform.