Liquid AI's LFM2.5-2.6B runs a tool-calling agent on-device
Trained on 34T tokens, the 2.6B model plans and calls tools at 220 tok/s in under 2.5GB — beating models 4x its size, with data that never leaves the device.
Copy markdown
A full agent that never phones home
LFM2.5-2.6B plans, calls tools, and runs multi-step tasks entirely on-device — phones, laptops, PCs, even robots. Nothing leaves the machine, and the marginal cost of each run is essentially zero.
220 tok/s on a Mac, under 2.5GB
It clocks 220 tok/s on an Apple M5 Max, 113 on a Ryzen AI Max+ 395, and ~30 tok/s on a phone — all in under 2.5GB of memory. On an H100 it pushes ~15K tok/s at high concurrency, roughly 1.3B tokens a day.
Beats models 4x its size on agentic work
Pre-trained on ~34T tokens with an agentic-RL finish, the 2.6B model tops every instruction-following benchmark (IFStruct 85.49, Multi-IF 80.07) and most tool-use ones (ToolSandbox 77.83) — outscoring models four times larger.
Grab it today: GGUF, MLX, vLLM, ONNX
Both the instruct and Base checkpoints are on Hugging Face now, with a 128K context window and runtimes for llama.cpp, MLX, vLLM, SGLang and ONNX. One caveat: it ships under Liquid's own LFM Open License v1.0 — free for small companies, not a standard permissive license.
Build this weekend: a private offline agent
With tool-calling at near-zero marginal cost and no data leaving the device, you can wire a fully offline assistant — file search, email triage, a robot controller — that runs on a laptop or phone with no API keys and no per-token bill.
Elsewhere: LLM 0.32 adds reasoning traces
Simon Willison's LLM CLI hit 0.32, adding captured reasoning traces, OpenAI Responses API support, server-side tools and smarter logging — useful for scripting and comparing models straight from the terminal.