MacPaw and Liquid AI Build a Private On-Device AI Stack Challenging Apple Intelligence

MacPaw and Liquid AI have announced a partnership to co-develop a full local AI stack for the Mac. The goal: bring private, fast, offline-capable AI to millions of Mac users without routing every query through a cloud server. The first product built on this stack is Eney, MacPaw's macOS AI assistant, with a production release planned for later this year.

Three layers, one stack

The collaboration combines Liquid Foundation Models (LFMs), optimized for macOS, with MacPaw's Elix inference engine and Mnemos memory system. The three components each handle a distinct job:

• LFMs — Liquid AI's foundation models, fine-tuned for macOS assistant tasks and designed to run efficiently on Apple silicon.
• Elix (Eney Local Intelligence MLX) — MacPaw's local engine that defaults to on-device processing and storage, keeping reasoning, context search, skill execution, and conversation history off the cloud.
• Mnemos — MacPaw's context system that continuously indexes your macOS environment to deliver personalized, situationally relevant responses.

Together, they let Eney run assistant and agentic workflows entirely offline, with personal data staying on the device.

How this differs from Apple Intelligence

Apple already ships local models to developers, so the distinction matters. Liquid AI CEO Ramin Hasani explained it to TechCrunch: "Before training our models, we select an architecture that is different and tailored to the hardware. That allows us to really have the most efficient version of intelligence that runs directly on the device, with benefits like privacy and security."

The deeper differentiator is adaptability. "We are also building a customization stack around models," Hasani said. "This means that with user input, the models can use the data and improve. We want our models to be adaptable and become more intelligent over time." Apple Intelligence models don't learn from personal usage patterns. Mnemos is specifically designed to do that, building a persistent, evolving picture of how a user works.

Why LFMs run well on-device

Liquid AI's models use a hybrid architecture combining convolutions and attention, rather than a standard transformer. That design trades theoretical generality for efficiency on constrained hardware, which matters when you're targeting a MacBook rather than a data center GPU.

The numbers back this up. The LFM2.5-2.6B model fits under 2.5 GB of memory and runs at 220 tokens per second on an Apple M5 Max, and 113 tokens per second on an AMD Ryzen AI Max+ 395. The LFM2-8B-A1B, an 8.3B mixture-of-experts model with only 1.5B active parameters, already runs locally on a MacBook Pro with an M2 Pro chip. At those speeds, a local assistant can handle real-time task execution without the lag that makes cloud-dependent tools feel slow.

Liquid AI has also released a range of model sizes: dense checkpoints from 350M to 2.6B parameters, and a series of LFM2-Nanos fine-tuned for specific tasks including tool and function calling, retrieval-augmented generation, and mathematical reasoning. The MacPaw-specific models will be fine-tuned on top of this foundation.

The Setapp angle is the real story

Eney is the launch vehicle, but MacPaw's longer-term plan is to open the local inference stack to developers building on Setapp, its subscription app platform. That changes the economics significantly for Mac indie developers.

Setapp has over 150,000 paying subscribers and more than 250 apps across productivity, creativity, and utility categories. MacPaw's 2024 revenue was $90 million, with 15% coming from Setapp. The company is also experimenting with credit-based pricing for AI operations, where users consume credits scaled to task complexity.

For developers, the proposition is concrete: instead of wiring up a cloud inference backend, paying per-token API bills, and handling privacy compliance independently, they could tap a shared local inference layer already running on their users' machines. If background agent loops run on the user's own device, hosted-API inference costs stop being a competitive moat and become an overhead line to justify.

Who's behind the deal

• Oleksandr Kosovan — CEO and founder of MacPaw, a Kyiv-founded company with offices in Boston and across the EU, operating since 2008.
• Ramin Hasani — Co-founder and CEO of Liquid AI, founded in 2023 alongside Daniela Rus and Alexander Amini.
• Liquid AI — An MIT spinout that raised $250M at a $2.35B valuation, with AMD leading its Series A. Existing customers include Mercedes-Benz, Insilico Medicine, and Shopify.

Who gains, who feels pressure

Privacy-conscious Mac users and indie developers in the Setapp ecosystem come out ahead. Users get an assistant that doesn't send every query to a remote server. Developers get an on-device inference primitive they don't have to build or pay for themselves.

The pressure lands on two groups. Cloud AI API providers face direct substitution if Setapp developers can route common tasks through local LFMs at no per-call cost. Apple faces a subtler challenge: a third-party stack, running on Apple silicon, positioning itself as the default local AI layer for macOS, the same territory Apple Intelligence occupies.

Liquid AI also gains something it hasn't had before: a high-visibility consumer deployment. The company has operated mostly in enterprise and industrial verticals. Getting LFMs into a product used by millions of Mac users is a concrete proof point for the "efficient models everywhere" thesis, and a useful signal for future enterprise deals.

What comes next

Eney ships later this year as the first product on the stack. After that, MacPaw plans to extend the shared infrastructure across its own product ecosystem before opening it to third-party Setapp developers. The timeline for that broader rollout hasn't been specified.

Financial terms were not disclosed. No equity exchange or acquisition is involved; this is a co-development agreement. MacPaw's CEO has framed it as a platform play rather than a single product launch.

For the on-device AI space, this partnership adds to a pattern that's becoming hard to ignore: the inference cost curve is dropping fast enough that running capable models locally is a product decision, not a research project. The open question is which stack ends up as the default substrate for local AI on the Mac — Apple's, Google's, or an independent layer like this one.