On August 4, 2026, Liquid AI announced the official release of the LFM2.5-2.6B model. This is an open-source model specifically designed for agent workflows, with a parameter scale of 2.6 billion. The model supports fully local operation on end-user devices such as smartphones, eliminating the need for reliance on cloud-based APIs. Data never leaves the device, and the marginal cost per operation is nearly zero. LFM2.5-2.6B achieves a decoding speed of 220 Tokens per second on the Apple M5 Max, demonstrating exceptional performance in instruction execution and tool usage tasks, surpassing multiple larger-scale models.