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

Liquid AI's new LFM2.5-2.6B beats models four times its size at tool-calling and runs entirely on-device — down to a Raspberry Pi. Here's what that… Aliteq Lena Fischer · AI & Local Compute Editor

LFM2.5-2.6B has 2.69B parameters, pretrained on ~34 trillion tokens, with a 131,072-token context window.

It beats the 9B-parameter Qwen3.5-9B on ToolSandbox, Multi-IF, and IFStruct despite being roughly a quarter of the size — but loses to it on BFCLv4 and coding benchmarks.

On an Apple M5 Max it runs at 220 tokens/second using under 2.5GB of memory; on an AMD Ryzen CPU, 113 tokens/second; on a phone, around 30 tokens/second.

It's built for tool-calling, data extraction, RAG, and long-running agent workflows — not for heavy coding or deep knowledge-recall tasks.

Weights are open on Hugging Face under Liquid's own lfm1.0 license, with day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX.

Yes, and it costs nothing to test — the weights are open and it runs on hardware you probably already own. Just be honest about which job you're hiring it for; it's a tool-calling specialist, not a…

Is this worth trying?

Read the full story

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