Liquid AI Releases LFM2.5-2.6B Model Optimized for Mobile and Edge Devices
Liquid AI has released LFM2.5-2.6B, a 2.69-billion parameter model featuring a 128K context window and native tool-calling support. The model is optimized for local execution and is already compatible with llama.cpp via the GGUF format. This release demonstrates that highly capable agentic models with large context windows can run efficiently on consumer hardware, achieving speeds of up to 30 tokens per second on mobile phones. It enables cost-effective, local deployment of AI agents for tasks like data extraction and file operations without relying on cloud APIs. While the model outperforms larger models like Qwen3.5-9B on benchmarks like ToolSandbox and IFBench, it remains weaker in coding and knowledge-heavy tasks. Additionally, running a 128K context window on mobile devices may face memory constraints due to KV cache overhead.
## BACKGROUND
Liquid Foundation Models (LFMs) are a class of generative AI models designed for high adaptability and efficiency, utilizing dynamic architectures that can modify computational graphs during inference. The GGUF format is a binary file format designed for fast loading and running of quantized models on consumer hardware, primarily used with llama.cpp, an open-source C/C++ inference engine.