Liquid AI's LFM2.5-2.6B runs a 9B-class agent on your phone
Open weights, under 2.5 GB, 220 tok/s on a laptop, and it edges models 4x its size on tool use. The license is the catch. Plus: Jeff Dean's new lab.
Copy markdown
A 2.6B model that punches at 9B
LFM2.5-2.6B scores 77.83 on the ToolSandbox agent benchmark, edging Qwen3.5-9B's 76.44 (a model nearly 4x larger) and leading every instruction-following test Liquid ran. Capable on-device agents just stopped needing a big model.
220 tok/s on a laptop, under 2.5 GB
It runs 220 tok/s on an Apple M5 Max, 113 tok/s on an AMD Ryzen AI Max+ 395, and about 30 tok/s on a phone, all in under 2.5 GB of RAM, with a 128K context and 16 languages. No GPU and no API key required.
Run it today, fully offline
Weights are on Hugging Face now with day-one support in llama.cpp, MLX, vLLM, SGLang, and ONNX, plus native function calling. This weekend you could stand up a private agent or RAG app whose data never leaves the device.
Read the license before you ship
It is open-weight but under Liquid's own LFM1.0 license, not Apache or MIT: commercial use is free only if your organization earns under $10M a year. Above that, you need a separate license from Liquid.
Elsewhere: Jeff Dean's landing spot is Discovery Loop
The Google veteran leaving after 27 years is co-founding Discovery Loop with Sanjay Ghemawat to automate scientific and engineering research, backed by Radical Ventures. No product yet, but it is where two of computing's most influential engineers are betting next.