Liquid AI's LFM2.5 230M targets phones and robots

Liquid AI has released LFM2.5 230M, a small language model aimed squarely at the edge. At roughly 230 million parameters, it is built to run directly on hardware where cloud inference is impractical or undesirable — phones, single-board computers like the Raspberry Pi, and robotics platforms, according to the company's announcement.

The pitch is straightforward: rather than routing every request to a remote API, developers can embed a capable text model into devices that need to work offline, respond with low latency, or keep data on-device for privacy reasons. That trade-off — accepting a smaller model in exchange for local execution — is increasingly the defining constraint of on-device AI.

Why it matters

Most attention in open models flows to systems with billions of parameters, but the sub-1B tier is where a lot of practical deployment happens. A 230M model is small enough to fit comfortably in memory on modest hardware while still handling structured text tasks.

• Runs on phones, Raspberry Pi–class boards, and robots
• Sits in the under-1-billion-parameter class for tight memory budgets
• Positioned for offline, low-latency, and privacy-sensitive use cases

Liquid AI has been building out its LFM2 line with an emphasis on efficient architectures, and this 230M entry extends that family toward the smallest end of the spectrum. Full licensing terms and specifics are outlined on the company's blog.