No cloud, no GPUs, no problem: Liquid AI's new model LFM2.5-2.6B brings powerful AI agents to devices as…

Liquid, the AI startup formed in 2023 by former MIT computer scientists, earlier this week debuted LFM2.5-2.6B, an open-weight language model built for agentic workloads. In release materials and a VentureBeat interview, the company’s researchers framed the model as one that can run entirely on local hardware—from smartphones and laptops down to a Raspberry Pi—without cloud inference or GPUs.

LFM2.5-2.6B is sized at 2.6B parameters and ships as open weights rather than a closed API. The product pitch centers on local inference for agent-style tasks: planning, tool use, and multi-step work that usually depends on remote servers. Liquid’s claim is that the same model can sit on phones, laptops, and single-board computers such as a Raspberry Pi, so agents do not need a network path to a GPU cluster to complete a turn.

For engineers building on-device agents, that matters because latency, privacy, and offline operation stop being deployment footnotes. An agent that runs next to sensors, files, or user data on the device can keep prompts and intermediate state local, avoid round-trips to a cloud endpoint, and still function when connectivity is poor or disallowed. The Raspberry Pi class of target also signals that the cost and power budget for an agent host can drop to hobbyist and edge hardware, not only phones and laptops with discrete GPUs.

The competitive frame is on-device and edge LLMs versus cloud-first agent stacks. Most production agents still call large hosted models; Liquid is pushing a smaller open-weight alternative aimed at agentic work on CPUs and consumer silicon. Open weights also change the integration path: teams can inspect, quantize, fine-tune, and embed the model inside their own runtimes instead of renting tokens through a provider API.

What to watch next is whether real agent pipelines—tool calling, long multi-step sessions, and recovery from bad intermediate steps—hold up on phones and Raspberry Pi–class boards at usable latency and memory use. Builders evaluating LFM2.5-2.6B should measure end-to-end agent loops on the exact devices they ship, not only chat-style single prompts, and treat the VentureBeat-backed local/no-GPU claim as a deployment target to verify against their own hardware and agent scaffolds.