A new AI model just ran real agent tasks on a Raspberry Pi — no GPU, no cloud bill, nothing

Liquid AI released a 2.69-billion-parameter model this week called LFM2.5-2.6B, and on its own tool-calling benchmark it beats Qwen3.5-9B — a model more than three times its size — 77.83 to 76.44 on ToolSandbox. It runs entirely on-device: phones, laptops, PCs, and, according to VentureBeat, hardware as modest as a Raspberry Pi. No GPU, no cloud API bill, no data leaving the device. That's the real claim, and the numbers behind it are worth walking through before you decide what it actually changes for you.

The numbers behind the claim

Read that table honestly and the picture is more interesting than a clean win. LFM2.5-2.6B leads on three of the four instruction-following and tool-use metrics against a model nearly four times larger — that's the genuinely impressive part, since parameter-efficient tool-calling is exactly what makes a model viable to run on a phone in the first place. But it loses on BFCLv4, a broader function-calling benchmark, and clearly loses on LiveCodeBench v6, a coding-specific test. This is a specialist model, not a smaller version of a generalist one. Liquid AI says as much directly: it's built for agentic workloads, not agentic coding.

How fast is it really, and where

What it's actually good for (and bad for)

• Good: tool calling and multi-step agent workflows — the exact thing it's benchmarked against.
• Good: structured data extraction and long-context document work, helped by the 131K context window.
• Good: retrieval-augmented generation (RAG) pipelines running entirely on-device.
• Good: always-on background automation — calendar management, workflow routines, connectivity-limited settings like vehicles and robotics.
• Bad: agentic coding — it loses clearly to larger models on LiveCodeBench v6.
• Bad: deep knowledge-heavy question answering, where raw parameter count still matters.

The cost angle is the part I think gets underplayed in coverage like this. Once a model this size is downloaded, the marginal cost of every additional run is effectively zero — no API metering, no per-token cloud bill, no data leaving the device. That's a genuinely different economic model from calling a hosted API for every agent step, and it's why Liquid AI is positioning this specifically at regulated industries and always-on background automation rather than at chatbot replacement. Weights are open on Hugging Face under Liquid's own lfm1.0 license, with day-one support across llama.cpp, MLX, vLLM, SGLang and ONNX — so there's no lock-in to a specific runtime to try it.

My take: this doesn't replace the GPU-buying conversation this site spends most of its time on — it narrows it. If what you actually want is a general local chatbot that can code, reason deeply, and hold a wide-ranging conversation, you still want real VRAM; see how much RAM you need for local AI and whether an 8GB card is enough for that path. But if what you actually need is a background agent that calls tools, extracts data, and runs a workflow without touching the cloud, LFM2.5-2.6B genuinely means you might not need to buy anything at all. Those are two different jobs, and conflating them is how you end up either overspending on hardware you didn't need or under-provisioning a project that actually needed a real GPU.

The bigger trend worth watching is how fast the on-device tier is closing the gap on tasks that used to require real infrastructure. A year ago, tool-calling agents this capable meant a cloud API call or a dedicated GPU box. Now it means a phone you already own. That doesn't kill the local-GPU market — coding, reasoning and general-purpose assistants are still firmly GPU territory — but it does mean the next question worth asking isn't just what to buy, but whether you need to buy anything at all.