Liquid AI has introduced a new on-device agentic model, LFM2.5-2.6B, designed to perform complex tasks such as planning and tool calling without the need for cloud-based processing. This model utilizes a combination of 22 short convolution blocks and 8 GQA blocks across 30 layers to handle large amounts of context and decode information at a rate of 220 tokens per second. The model's open weights are available in various formats, including GGUF, MLX, and ONNX. This development is significant because it enables more efficient and localized processing of complex tasks, which could have implications for various industries and applications.