Liquid AI's 2.6B model runs agents on a laptop with no cloud bill

The cloud AI bill has been the price of admission for building anything smart. Send the request out, wait for the server, pay per token, hope the user has signal. That deal is starting to break. Small models now run on the phone and laptop in your pocket, fast enough to feel real, and a few companies are turning that into a product bet. Let me catch you up.

The deep cut

• On-device shifts the cost from you to the chip. LFM2.5-2.6B runs agents with no cloud inference bill on a laptop.
• Local means it works when the signal dies. A CarPlay assistant goes silent in a tunnel because the model lives on a distant server.
• The model is rarely your slowest part. One builder found input handling, not inference, was screaming on low-end phones.

Small models that hold their own

Liquid AI's LFM2.5-2.6B is a 2.6 billion parameter model that runs in under 2.5 GB of memory. It hits 220 tokens per second on an Apple M5 Max and 113 on a Ryzen CPU. At 30 tokens per second, it runs on a phone. Small enough for everyday hardware, fast enough to feel responsive.

The pitch is not just size. On tool use and instruction following, it beats models up to four times larger. It tops every instruction-following benchmark in the group and most tool-use ones. Coding is where bigger models still win clearly, so the team says reach for something bigger there. That honesty matters. This is a specialist, not a do-everything brain.

Betting the store on local

MacPaw is building this into a business. Ivan Mehta reported that the company partnered with Liquid AI to build an on-device inference system called Elix, plus a local memory system, for its assistant Eney. The goal is to hand that stack to developers on SetApp, its app store with over 150,000 paying users.

The selling points are privacy, offline use, and no per-request cloud bill. Liquid AI's Ramin Hasani said they pick a model architecture "tailored to the hardware" to get "the most efficient version of intelligence that runs directly on the device." MacPaw is testing credit-based pricing, where a user's credits cover a set number of AI operations. That is a real answer to the question every AI product owner faces: who pays for the compute, and how.

Silence in the tunnel

Apple opened CarPlay to third-party voice AI apps in 2026. Sounds like a win. But as one build-along guide points out, those apps are front ends to cloud models. Drive into a tunnel, lose signal, and the assistant goes quiet. "The assistant was never really in your car," the author writes. Your dashboard was just a microphone.

The fix is a small model, speech-to-text, and voice all running on the box in the dashboard. No connection, no wait, nothing leaves the car. The build uses a 3-billion parameter model as the "sweet spot" for fast spoken replies. Local is not only about privacy. It is about the product still working when the network does not.

Profiling the wrong thing

Small and local does not mean easy. One builder spent days tuning prompts and quantized models, sure inference time was the problem. It wasn't close. Input processing was the bottleneck on low-end phones. He was optimizing the wrong number.

You see the same lesson at the chip level. In a study of five ways to compute an STFT on a Snapdragon SoC, moving work off the CPU to a low-power DSP or the NPU actually used more energy per chunk, by nearly two orders of magnitude. The offload's dispatch cost dominated a tiny operation. The takeaway for your team: measure the whole pipeline, not the piece you assume is slow. On-device performance is a systems problem, and your instincts will lie to you.

Three questions for your team

• For our next AI feature, what breaks when the user loses signal, and is that a bug or a dealbreaker?
• If we move inference on-device, who was paying the cloud bill, and does the chip cost land on us or the customer?
• Before we spend a sprint optimizing the model, have we actually profiled the full pipeline to confirm the model is the slow part?