Liquid AI releases LFM2.5-2.6B, an on-device agentic model that runs on phones and laptops

Liquid AI released LFM2.5-2.6B on August 4, an open-weight, 2.6-billion-parameter model built specifically to run local agentic workloads entirely on-device, rather than routing tool calls and planning steps through a cloud API.

What's new

The model is small enough to fit on consumer hardware with a memory footprint under 2.5 GB, and Liquid AI reports CPU inference speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen chip, with GPU throughput reaching roughly 15,000 output tokens per second at high concurrency. It supports a 128K-token context window. In Liquid AI's own words: "LFM2.5-2.6B is built to power capable agents entirely on-device. It supports tool calling and multi-step workflows while staying small and fast enough for everyday hardware."

The company is positioning the model specifically around agentic capability rather than general chat quality, claiming it is "competitive with models 4x larger on tool use, instruction following, and multi-step agentic tasks." The release ships as open weights on Hugging Face, alongside a base variant for further fine-tuning.

Context

LFM2.5-2.6B extends Liquid AI's line of small, efficiency-focused models — the company shipped the encoder-focused LFM2.5-Encoders release earlier — built around the thesis that a large share of agentic workloads (tool calling, multi-step task execution, simple planning loops) don't require frontier-scale models if the smaller model is trained specifically for that structure. The pitch is direct: "This enables developers to deploy agents everywhere, keep data private on the device, and scale usage without a cloud inference bill."

Why it matters

Most agentic AI today depends on a round trip to a hosted frontier model, which means latency, per-token cost, and a hard requirement for network connectivity. A model that can plan and call tools locally, fast enough to feel responsive on a laptop or phone CPU, changes the economics and privacy profile for a real category of applications — offline assistants, on-device automation, and privacy-sensitive workflows where data can't leave the device. Whether LFM2.5-2.6B's benchmark claims hold up against larger cloud models in practice is unverified outside Liquid AI's own testing, but the release adds to a growing body of small, purpose-built agentic models competing on task-specific capability rather than raw parameter count.