Liquid AI puts a 2.6 billion parameter agent in a Raspberry Pi

Liquid AI has released LFM2.5-2.6B, an open-weight language model built to run AI agents directly on local hardware, including laptops, smartphones and Raspberry Pi computers. Carl Franzen reports for VentureBeat that the model targets companies that need lower latency, tighter data control or lower operating costs than cloud-based AI can offer.

The 2.6 billion parameter model supports a context window of up to 128,000 tokens and has native tool calling. This lets it interact with software tools, such as calendars, document systems and workflow applications. Liquid positions it for recurring, clearly defined tasks rather than complex coding or broad general-purpose reasoning.

According to company measurements, LFM2.5-2.6B uses under 2.5 GB of memory. Liquid reports about 30 output tokens per second on a smartphone, while its demonstrations show the model running on a Raspberry Pi. The figures have not been independently verified.

Designed for agent workflows

Liquid says it trained the model in agent environments including Hermes Agent and OpenClaw. Rather than optimizing only for chatbot conversations, the training focused on using tools, managing documents and completing productivity tasks. The company also built a phone-native agent harness that can run background routines locally.

The release is available through Hugging Face, alongside a base model for fine-tuning and the company’s LEAP fine-tuning framework. It works with inference tools including llama.cpp, MLX, vLLM, SGLang and ONNX.

Liquid’s own benchmark results show strong instruction-following and tool-use performance against several larger small models from Google and Alibaba. Those claims should be treated cautiously because they come from the vendor. The model is text-only, unlike some multimodal alternatives.

Its license may matter for business users. The LFM Open License allows commercial use for organisations with annual revenue below $10 million. Larger companies need a separate agreement with Liquid AI, unlike models released under permissive Apache 2.0 or MIT licences.