Liquid AI's LFM2.5-2.6B runs a real agent on a phone

Liquid AI released LFM2.5-2.6B on August 4 — an open-weight model built to plan, call tools, and grind through multi-step tasks without a cloud API in the loop. It fits in under 2.5 GB and decodes at 220 tokens/s on an M5 Max. For anyone who has watched a per-token bill scale with usage, that combination is the story.

What actually happened

Per Liquid AI's release post, the model is 2.6 billion parameters, pre-trained on roughly 34 trillion tokens, and post-trained specifically inside agent harnesses rather than on generic chat data.

The reported numbers: 77.83 on ToolSandbox, 59.17 on IFBench, 56.88 on BFCLv4, 59.41 on LiveCodeBench v6, 51.87 on AIME25. Liquid claims it leads its size class on instruction-following and most tool-use tests, and trades punches with models several times bigger.

Speed, by hardware: 220 tokens/s decode on an M5 Max, 113 tokens/s on a Ryzen AI Max+ 395, about 30 tokens/s on a phone. On an H100 at high concurrency, close to 15,000 output tokens/s.

Weights are on Hugging Face — base and post-trained — with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX across Apple Silicon, AMD, Qualcomm, and NVIDIA.

The benchmarks are vendor-reported. We'd run your own workload before believing any of them. But the runtime support list is verifiable, and it's the part that determines whether you can actually ship this.

Why on-device AI agents matter for your business

A 2.6B model that does tool calls changes which features you can afford. Most of the "AI" in a small business app is classification, extraction, routing, and filling in a form — work that never needed a frontier model. Those calls are exactly where per-token pricing quietly compounds. Move them to a local model and the marginal cost of the tenth thousand run equals the first.

Thirty tokens per second on a phone is slow for chat and fine for a background job. Don't judge the speed against a streaming chat window. Judge it against the task: a nightly categorization pass, an offline intake form that parses itself, a field tech's app that works in a basement with no signal. That work doesn't need to feel instant. It needs to finish and not cost anything.

Open weights with no restrictions is a portability guarantee you can hold. A file on disk cannot be deprecated, rate-limited, or repriced mid-quarter. We've written enough about retired model aliases and vanished API endpoints to say plainly: keeping one capable local model in your stack is cheap insurance, even if it's the fallback and not the default. Pair it with the on-device runtime work happening at the desktop layer and the pattern gets easier to ship every month.

Sources: Liquid AI, Liquid AI on X.