Wispr Flow - Don’t type, just speak
Mini toolsLFM2.5-230M
By Liquid AI
Model family: LFM
LFM2.5-230M is Liquid AI's smallest model, pre-trained on 19T tokens with a 32K context extension phase. Post-training uses a three-stage recipe: supervised fine-tuning with distillation from LFM2.5-350M, direct preference optimization, and multi-domain reinforcement learning. It excels at tool use (BFCLv3/v4) and data extraction (CaseReportBench), competing with models more than twice its size on instruction following and applied tasks. Inference is supported across llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX from day one. CPU performance reaches 42 tok/s on Raspberry Pi 5 and 213 tok/s on a Snapdragon Gen4 device. The model has been demonstrated running on-device on a Unitree G1 humanoid robot for natural-language skill sequencing. Base and post-trained checkpoints are available on Hugging Face under an open-weight license. Not recommended for reasoning-heavy tasks such as advanced math, code generation, or creative writing.
New
Text
Gen 7
Overview
LFM2.5-230M is a 230M-parameter lightweight language model built on the LFM2 architecture, designed for fast inference across edge and cloud environments. Pre-trained on 19T tokens with a 32K context window, it targets tool use, data extraction, and agentic workflows. Runs on CPUs including Raspberry Pi 5 and mobile SoCs. Not recommended for math, code generation, or creative writing.
About Liquid AI
Liquid AI is an MIT spin-off building efficient general-purpose AI models (Liquid Foundation Models, or LFMs) that run on edge devices with less memory and power.
They recently raised $250M in Series A funding to scale model development and deployment.
Last updated: June 26, 2026
