LFM2.5-2.6B arrives for local agent deployment on edge devices
Drafted by our automated systems from the sources listed at the end of this article, and reviewed before publication. Every fact here should be traceable to one of those links; we publish no benchmark, price or figure that a linked source does not state. Where a company makes a claim about its own product, we report it as their claim.
According to the Hugging Face blog post, LFM2.5-2.6B is designed for local agent deployment on edge devices and comes pre-trained on approximately 34T tokens. Its context window has been extended to 128K through a mid-training phase. The blog states that post-training turns the base model into an agent via four stages, including supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning.
Inference speed and ecosystem support
Liquid AI reports decode speeds of 220 tokens/s on an Apple M5 Max and 113 tokens/s on an AMD Ryzen CPU. The model ships with day-one support across several inference ecosystems, which include llama.cpp, MLX, vLLM, SGLang, and ONNX according to the blog post.
Is it worth it?
This release is worth trying if you need on-device agents for instruction following and tool use, though you should look elsewhere for coding tasks. A reader would want to check performance against their own specific local hardware constraints and agent harnesses before switching production workloads.
What this changes for tools we list
- Hugging Face — The LFM2.5-2.6B and LFM2.5-2.6B-Base models are now available to download directly on Hugging Face.
Sources
- Deploy local agents everywhere with LFM2.5-2.6B — Hugging Face, 4 August 2026