Liquid AI just put a working AI agent on your phone

What happened

Liquid AI, the MIT-spinoff behind the LFM family of efficient models, shipped LFM2.5-2.6B on August 4. It is a 2.6-billion-parameter model trained on roughly 34 trillion tokens, and it arrives in both base and instruction-tuned versions on Hugging Face, with day-one support across the major local runtimes (llama.cpp, MLX, vLLM, SGLang, ONNX).

The pitch is unusually specific: this is not a general chat model shrunk down, but a model post-trained end-to-end for agent work. Liquid AI says it went through four stages — supervised fine-tuning on agentic data, distillation from domain-specific teacher models, a multi-domain on-policy distillation step, and reinforcement learning inside live agent harnesses rather than static benchmarks.

The company announced the release on its blog and on X: "It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots. Data never leaves the device, and the marginal cost of each run is essentially zero."

Why it matters

Agents have been the one workload where small models consistently fell apart — a model that fumbles one tool call in a five-step chain fails the whole task. Liquid AI is claiming, with benchmark numbers, that a model small enough to run in under 2.5 GB of memory can now hold its own against models several times its size on exactly that workload.

If the numbers hold up in practice, the economics change for anyone building agent features. Running each step through a cloud model meters every user action; a competent on-device agent makes each run effectively free, works offline, and keeps user data on the device. For teams shipping mobile or desktop apps, that is the difference between an agent feature you ration and one you turn on by default.

Key details

On tool use, Liquid AI reports its 2.6-billion-parameter model edging out Qwen3.5-9B — a model more than three times larger — on the ToolSandbox agent benchmark, and beating comparable small rivals from Google and Alibaba by wide margins on instruction following. These figures are self-reported by Liquid AI and have not yet been independently reproduced; treat them as the company's own numbers.

Speed is the more concrete claim: around 220 tokens per second on a current MacBook chip, and about 30 tokens per second on a phone — fast enough for interactive agent loops, not just batch work. The model handles long contexts (128K tokens), which matters for agents that accumulate tool results across many steps.

One caveat for commercial teams: the weights are open but ship under Liquid AI's own LFM license, which is free for smaller companies but requires a commercial arrangement above roughly $10 million in annual revenue. Read the terms before you build on it.

Bottom line

If you are building agent features and have been defaulting to cloud models for reliability, LFM2.5-2.6B is the first small model this year that is worth an afternoon of testing on your actual tool-calling workload — the license is the only fine print, and the download is free.

Sources

Official announcement — Liquid AI blog · Release post — Liquid AI on Hugging Face · Model weights — Hugging Face · Announcement on X · Coverage — The University Network. Verified: release date (August 4, 2026), model size, training scale, context length, memory footprint, speed figures and benchmark scores confirmed against Liquid AI's official Hugging Face post and cross-checked against three independent write-ups. Benchmark scores are Liquid AI's own reported numbers and are labelled as such.