🧾
2 entries
Title: LFM2.5-230M
Entries are written by users and may reflect personal opinions or experiences.
Entries
A brilliant micro model that can run on a raspberry pi, Jetson orin or smartphone locally. they tried the model on an Unitree G1 robot, running locally.
Also supports mlx, llama.cp, vLLM and ONNX. Can run on pretty much every where. Even without a GPU.
Tailored for agentic workloads, data extraction and tool use. Can be a wise choice for data extraction, skill selection, command decomposition and the like. One of the smallest models that you can actually make use of.
A must try.
Liquid AI has released LFM2.5-230M, its smallest model to date, a 230-million-parameter model built on the company’s LFM2 architecture and explicitly optimized for fast, on-device agentic workloads across phones, robots, and edge hardware.Pre-trained on 19 trillion tokens with a 32K context window and further refined via distillation from the larger LFM2.5-350M, the model delivers strong performance on instruction following, data extraction, and tool use—often matching or beating models more than twice its size—while achieving impressive speeds: 213 tokens per second on a Galaxy S25 Ultra CPU and 42 tokens per second on a Raspberry Pi 5. It runs efficiently on CPUs, NPUs, and GPUs via llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX, and Liquid AI demonstrated it running entirely on-device on a Unitree G1 humanoid robot’s NVIDIA Jetson Orin to turn natural-language commands into structured multi-step plans. The release positions the model for both large-scale data pipelines and lightweight, privacy-focused agentic applications at the edge.