- MacPaw partnered with Liquid AI to run its Eney assistant fully on the device — no cloud round-trip, no per-token bill.
- Liquid AI is building an on-device inference engine called Elix plus a local memory system; its LFM2 models run up to 2x faster on CPU than comparable models.
- Apple’s Foundation Models framework already ships free on-device LLM inference to every iOS 26 app, and now opens to Claude and Gemini through one interface.
- The center of gravity for cost, privacy, and latency is moving from the data center to the chip in your pocket.
“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.” That line from Liquid AI CEO Ramin Hasani, given to TechCrunch on August 5, 2026, is the quiet thesis behind one of the year’s most underrated shifts: the smartest place to run AI may no longer be a distant data center, but the device already in your hand.
MacPaw Just Put a Full Assistant Where the Cloud Can’t Reach
What MacPaw and Liquid AI Are Actually Building
Ukraine-based MacPaw — the studio behind CleanMyMac and the SetApp subscription store — has teamed up with Liquid AI to power its products with locally hosted AI models, and eventually to hand that stack to outside developers. The first target is Eney, the AI assistant MacPaw unveiled last year. Instead of a cloud-only assistant, the company is building a locally hosted version, and it has brought in Liquid AI to develop an on-device inference system called Elix along with a local memory system.
The payoff, according to MacPaw CEO Oleksandr Kosovan, is that locally hosted models let users run assistants and agentic workflows offline — no signal required. MacPaw is reworking SetApp, which already has more than 150,000 paying users, around AI apps with credit-based pricing, where each AI operation draws down credits based on task complexity. The platform will still route to cloud models from the likes of Google when needed, positioning itself as a one-stop shop rather than a walled garden.
Trend Insight — The interesting part is not one assistant going local. It is that a distribution channel with a six-figure paying audience is betting its economics on on-device inference as the default, with the cloud demoted to a fallback.
Why ‘Small’ Models Are Suddenly Beating Big Ones
Liquid AI’s Efficiency Bet
Liquid AI, an MIT spinout that raised $250 million in December 2024 in a round led by AMD, brands itself around “device-native” foundation models. Its LFM2 family spans roughly 350M to 3B parameters and leans on a non-Transformer design — gated short convolutions paired with a small number of grouped-query-attention blocks. The company reports up to 2x faster prefill and decode on CPUs versus similarly sized models, and about 3x better training efficiency than its previous generation.
The efficiency story gets sharper at the bottom of the range. Liquid AI’s smallest model, LFM2.5-230M, reportedly outperforms models four times its size at data extraction while being small enough to run virtually anywhere. Hasani frames the advantage as customization, not just size: “We are also building a customization stack around models. 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.”
Trend Insight — For years the axiom was “bigger is better.” The device era flips the metric: what matters is intelligence per watt and per millisecond, not the top line on a leaderboard a phone can never load.
Apple Already Made Offline AI Free — and That Resets the Math
The Economics Developers Actually Feel
MacPaw and Liquid AI are not moving into empty territory. With iOS 26, iPadOS 26, and macOS 26, Apple turned its on-device model — about 3 billion parameters — into a first-class developer tool through the Foundation Models framework. Any app can call it for local LLM inference with no API costs, no network dependency, and privacy by default, because the data never leaves the device.
At WWDC 2026, Apple went further and opened the framework to any provider: a single Swift interface now reaches Apple’s local model plus cloud models like Anthropic’s Claude and Google’s Gemini. The practical result is that the everyday AI features — summarizing, classifying, extracting, answering — can ship for free on the user’s own hardware. That quietly undercuts the server-plus-API-contract pattern that defined how AI apps were built from 2023 through 2025.
Trend Insight — When the marginal cost of an inference falls toward zero and the data stays on the device, the build-versus-buy calculus for AI features resets. Expect “local-first, cloud-optional” to become a product default rather than a premium privacy checkbox.
What This Means for Builders and Buyers
The Trade-offs Nobody Is Hiding
On-device is not a silver bullet. Frontier reasoning, very long context, and the heaviest multimodal jobs still favor the cloud, where memory and accelerators are effectively unlimited. But a growing share of everyday work — drafting, tagging, routing, private question-answering, and offline agents — fits comfortably inside a 1B to 3B model on modern silicon.
The pattern taking shape across 2026 is a hybrid one: a small local model handles the common case instantly and privately, then escalates to a cloud model only when the task genuinely demands it. MacPaw’s credit-based store and Apple’s provider-agnostic framework are two expressions of the same architecture. For teams shipping AI features, the strategic question is shifting from which model is smartest to how much can be kept on the device before a token meter ever starts running.
Trend Insight — The winners of the on-device wave will not be whoever trains the largest model. They will be whoever squeezes the most useful work out of the smallest model a customer already owns.
Related
- Open-Weight AI Caught the Frontier – But the Safety Gap Remains
- Anthropic’s AI Breached 3 Companies in a Controlled Cyber Test
- Autodesk’s $3.6B Bet to Own Operations AI
- A Benioff-Backed Startup Wants to Fix AI Deployment
- AI Biz Tech Digest – August 5, 2026
Sources
- TechCrunch – MacPaw taps Liquid AI to offer on-device inference to developers (Aug 5, 2026)
- Liquid AI Blog – Introducing LFM2.5: The Next Generation of On-Device AI
- VentureBeat – Liquid AI’s LFM2.5-230M beats models 4x its size at data extraction
- Apple Machine Learning Research – Apple’s On-Device and Server Foundation Models
AI Biz Insider · AI Trends EN · aibizinsider.com
댓글 남기기