MacPaw and Liquid AI push inference onto the Mac, not the cloud
MacPaw is building on-device inference with Liquid AI and plans to hand it to Setapp developers. The credit meter is the part worth reading.
MacPaw announced a long-term partnership with Liquid AI to build local AI on the Mac — and then said the quiet part: once the architecture settles, it plans to hand that on-device inference stack to the developers building apps for Setapp. An app marketplace turning itself into an inference distribution layer is a more interesting move than another assistant.
What actually happened
Per TechCrunch and MacPaw's own announcement, the deal pairs Liquid AI's Liquid Foundation Models — architected for the hardware rather than shrunk down to fit it — with two MacPaw components: Elix, its on-device inference runtime, and Mnemos, a local memory system.
The first target is a locally hosted version of Eney, MacPaw's assistant, capable of running offline. The second target is the one that matters: making the same stack available to third-party developers shipping through Setapp, which TechCrunch reports has north of 150,000 paying users. The platform will also broker cloud models from providers like Google, so developers get a single surface for both.
No terms, no launch date, no benchmark numbers disclosed. Treat the timeline as aspirational.
The detail we'd underline: Setapp is experimenting with credit-based pricing, where a user gets a number of AI operations scaled by task complexity. Local inference, metered.
Why on-device inference matters for your business
Marginal cost is what kills AI features, not capability. Most small software businesses can build the feature. What they can't model is the per-user monthly bill when 8% of users hammer it. Moving inference to the customer's silicon takes that variable to roughly zero and replaces it with a fixed hardware floor — you need a machine that can hold the model. That's a trade a lot of desktop products should be making and aren't.
Metered "local" AI is still metered. Read the credit line again. Even when the compute runs on the user's laptop, the platform that distributes the runtime can price the operations. If you ship through someone else's marketplace, you inherit their meter. Know the rate card before you design the feature around it.
Portability is the whole point of a local runtime — don't trade it away. A model file on disk is the most portable AI you can ship. Bind your app to a proprietary inference SDK and you've reintroduced the vendor dependency you moved local to escape. Keep the model interface behind your own abstraction, the same discipline we apply to any model layer.
Key takeaways
- MacPaw and Liquid AI are co-developing a local AI stack for macOS: LFM models plus Elix inference and Mnemos memory
- MacPaw plans to open the stack to Setapp developers — the marketplace becomes an inference channel
- Setapp is testing credit-based pricing where AI operations are metered by task complexity
- No terms, launch date, or performance benchmarks were disclosed
- On-device inference trades per-query cost for a hardware floor — often the right trade for desktop software
- Keep the model interface behind your own abstraction so a local runtime doesn't become new lock-in
The AI feature that ships is the one whose unit economics you can defend. We build model layers that route between local and hosted inference on cost and latency, so you can change your mind later without a rewrite. Run the numbers on your use case, or talk it through with us.
Sources: TechCrunch, MacPaw.
- #on-device-ai
- #local-inference
- #macpaw
- #ai-costs
- #developer-tools
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Microsoft's Orchard is open — read the agent pass rates, not the pitch
Microsoft Research's Orchard framework is MIT-licensed and public. Its own benchmarks show agents at 69.7% on SWE-bench and 59.6% on assistant tasks. Plan for that.
Read itGenOffice is open source. The AI inside it isn't.
Genspark open-sourced GenOffice, an Apache 2.0 AI office suite for Mac and Windows. Read the routing: model calls go through Genspark's servers, not your key.
Read it