Liquid AI runs a 2.6B agent model on phones

Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter agent model that the company says decodes at 30 tokens per second on a phone, 113 on a Ryzen AI Max+ 395, and 220 on an Apple M5 Max while using under 2.5 GB of memory. Liquid AI’s model announcement says the open-weight checkpoint supports tool calls and multi-step workflows, while its 128K context and training across roughly 34 trillion tokens are documented in the Hugging Face model card. Builders testing the economics of local inference on Apple hardware should file this as a small-model candidate for private, high-volume edge agents—not coding-heavy work, which Liquid says remains a weakness.