The very small model that wants to run everywhere

Liquid AI launches LFM2.5-230M, a 230-million-parameter model optimized for local agentic use on hardware like the Galaxy S25 Ultra and Raspberry Pi 5.

With LFM2.5-230M, Liquid AI pushes its LFM2.5 family into a tiny format, designed for fast inference, fine-tuning, and local agentic uses. The model has 230M parameters, was pre-trained on 19T tokens with a 32K context window, then post-trained via distillation from LFM2.5-350M, preference optimization, and multi-domain reinforcement learning.

Liquid AI primarily highlights its ability to run on highly accessible hardware: 213 tokens per second in decoding on a Galaxy S25 Ultra, 42 tokens per second on a Raspberry Pi 5, and out-of-the-box compatibility with llama.cpp, MLX, vLLM, SGLang, and ONNX. The model is offered in base and post-trained versions on Hugging Face, as open-weight. In terms of applications, Liquid AI positions it for data extraction, tool use, and lightweight embedded agents.

The company also demonstrates a test on a Unitree G1 humanoid robot, where the model serves as a skill selection layer from natural language instructions. At this size, Liquid AI, however, advises against heavy reasoning tasks, advanced coding, or creative writing.