Liquid AI launches LFM2.5‑2.6B edge AI model for smartphone deployment

Artificial‑intelligence startup Liquid AI announced the open‑source LFM2.5‑2.6B model, a 2.6 billion‑parameter agent‑compatible model designed for on‑device execution. Trained on a 34‑trillion‑token dataset and equipped with an extended tokenizer, the model supports non‑Latin scripts and can run locally on smartphones without cloud APIs.

In benchmark tests against larger models such as Gemma‑4 and Qwen‑3, LFM2.5‑2.6B led most instruction‑following and tool‑use tasks, matching or surpassing the performance of bigger competitors in agent workloads and STEM evaluations. The model processes up to 30 tokens per second on a phone, uses under 2.5 GB of memory, and can decode 220 tokens per second on an Apple M5 Max GPU, enabling private, low‑latency AI agents on consumer hardware.