LFM2.5 is a family of AI models designed specifically for on-device deployment at the edge. The product consists of the LFM2.5-1.2B model family, which represents Liquid AI's most capable release yet for edge AI deployment. It builds on the LFM2 device-optimized architecture and is optimized for instruction following capabilities to serve as the building block of on-device agentic AI.

The LFM2.5 family includes multiple specialized models: Text models offering uncompromised quality for high-performance on-device workflows, an Audio model that is 8x faster than its predecessor running natively on constrained hardware, and a Vision-Language Model (VLM) that boosts multi-image, multilingual vision understanding and instruction following. The family comprises Base, Instruct, Japanese, Vision-Language, and Audio-Language variants, with extended pretraining from 10T to 28T tokens and significantly scaled up post-training pipeline with reinforcement learning.

The models feature a hybrid architecture that enables extremely fast inference speed on CPUs with a low memory profile compared to similar-sized models. The Audio-Language model processes audio natively rather than using pipelined approaches, eliminating information barriers between components and dramatically reducing end-to-end latency. It includes a custom LFM-based audio detokenizer with significantly reduced latency that is 8x faster than the previous generation.

LFM2.5 enables access to private, fast, and always-on intelligence on any device, making it suitable for on-device use cases like local copilots, in-car assistants, and local productivity workflows. The models unlock new deployment scenarios across various devices including vehicles, mobile devices, laptops, IoT devices, and embedded systems.

The models are open-weight and available through multiple deployment frameworks including LEAP (Liquid's Edge AI Platform), llama.cpp for CPU inference, MLX for Apple Silicon, vLLM for GPU-accelerated serving, and ONNX for cross-platform inference. They support both CPU and GPU acceleration across Apple, AMD, Qualcomm, and Nvidia hardware, with optimized models available through partnerships with AMD and Nexa AI for NPU deployment.