Liquid AI's LFM2.5-2.6B Model Runs Powerful AI Agents on CPUs and Raspberry Pi

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

Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter open-weight language model with a 128,000-token context window and native tool calling, according to VentureBeat. The model is purpose-built for agentic workloads and engineered to run on CPU-only hardware — from laptops and smartphones down to a Raspberry Pi — without invoking cloud inference or dedicated GPUs.

Architecture and deployment surface

The LFM2.5 generation ships with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX, with both the post-trained checkpoint and a base variant (LFM2.5-2.6B-Base) available on Hugging Face. The underlying LFM2 architecture was designed around real-world CPU performance rather than GPU benchmarks, as Maxime Labonne, Liquid AI's head of post-training, stated in a post-launch interview. Company-reported measurements place decoding throughput at approximately 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, while holding memory consumption below 2.5 GB. On a smartphone, throughput falls to roughly 30 tokens per second. Fine-tuning is supported through LEAP, an open-source framework from the same team.

Target workloads and trade-offs

LFM2.5-2.6B is positioned for high-volume, well-defined agentic tasks — tool calling, document management, calendar and workflow automation, and always-on background routines — and for connectivity-limited environments such as vehicles and robotics. Coding-heavy workloads are explicitly deferred to larger frontier models; the company frames the model as complementary to cloud-grade systems rather than a replacement. The release carries a custom open-weight license, distinct from the permissive licenses in widespread use, which enterprise legal teams will need to review separately. The model is accessible through Apollo, Liquid AI's mobile app, for hands-on testing on consumer devices.

Verification checklist

The headline throughput numbers are vendor-reported and unaudited against independent benchmarks. Energy and latency claims on a Raspberry Pi in particular should be replicated on the exact hardware revision a deployment will target — throughput on a Raspberry Pi 5 does not translate directly to a Pi 4. The license terms remain the gating factor for regulated industries. The CPU-first design choice mirrors a broader pattern in distributed systems where deployment footprint and cost dominate the design calculus; comparable shifts are visible in multi-asset trading infrastructure rollouts targeting wider geographic reach.