Liquid AI's LFM2.5-2.6B runs a tool-calling agent on a Raspberry Pi
Open weights, 128K context and function calling in under 2.5 GB; the 2.6B model beats 8-9B rivals at instruction-following, no GPU. Plus a new LLM CLI.
Copy markdown
A 2.6B agent that fits on a phone
LFM2.5-2.6B decodes at 220 tokens/s on an Apple M5 Max and 113 tok/s on a Ryzen AI Max+, all inside 2.5 GB of memory. That's enough to run tool-using agents fully offline on a laptop, a handset, or even a Raspberry Pi — no cloud call, no GPU.
It punches two weight classes up
On instruction-following and tool benchmarks it tops far larger models: IFBench 59.17 vs Gemma-4 8B's 39.24, Multi-IF 80.07 vs Qwen3.5-9B's 62.55, ToolSandbox 77.83. The catch — it's weak on coding and knowledge-heavy trivia, so aim it at agents, data extraction, and RAG, not your codegen.
Open weights, day-one llama.cpp and MLX
Base and instruct checkpoints are on Hugging Face with a 131K-token context and GGUF, MLX, vLLM, SGLang, and ONNX support out of the box. The LFM Open License v1.0 is free for nonprofits and companies under $10M revenue — enough to ship a private, offline agent this weekend.
Elsewhere: LLM 0.32 adds reasoning traces
Simon Willison's LLM CLI and Python library shipped 0.32 with reasoning-trace capture, OpenAI Responses API support, server-side tools, and smarter logging — a fast upgrade if you script across many models from the terminal.