LFM2.5-2.6B

LFM2.5-2.6B is a 2.6 billion parameter dense language model developed by Liquid AI, specifically optimized for on-device agentic workloads. Released in August 2026, it belongs to the LFM2.5 family of hybrid models designed to balance high performance with the low-latency and memory requirements of edge deployment. The model is post-trained using reinforcement learning within agent harnesses to improve its reliability in executing multi-step workflows and native tool calling.

The model utilizes a unique hybrid architecture consisting of 30 layers: 22 double-gated short convolution blocks paired with 8 grouped-query attention (GQA) layers. This design allows it to process data more efficiently than traditional transformer-only architectures, particularly on hardware with limited resources. It was pre-trained on approximately 34 trillion tokens, resulting in a compact reasoning engine that competes with models significantly larger in size on specific instruction-following and tool-use benchmarks.

Key Capabilities

LFM2.5-2.6B is engineered for agentic tasks, such as data extraction, Retrieval-Augmented Generation (RAG), and multi-turn conversations. It features a 128,000-token context window, enabling the processing of long documents and extensive tool traces. The model supports 16 languages, including English, Chinese, French, German, Japanese, and Korean, making it versatile for global local-first applications.

While the model excels at structured output and tool orchestration, Liquid AI recommends against using it for knowledge-heavy tasks or complex agentic coding, where larger models remain more reliable. Its primary strength lies in high-throughput local execution, achieving notable performance on consumer-grade hardware such as Apple Silicon and AMD Ryzen CPUs, typically operating in under 2.5 GB of memory.