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

Liquid AI has released LFM2.5-230M, a compact language model with roughly 230 million parameters designed to run directly on constrained hardware rather than in the cloud. The company frames it for phones, Raspberry Pi boards, and robotics, where memory and power budgets rule out larger models.

At this size, the appeal is less about topping leaderboards and more about placement. A model small enough to fit comfortably on-device can respond without a network round trip, keep data local, and operate in environments where connectivity is unreliable — the exact conditions faced by embedded systems and mobile applications.

Why it matters

The industry's attention tends to follow ever-larger frontier models, but a parallel push toward sub-billion-parameter systems is opening up new deployment targets. LFM2.5-230M is squarely part of that trend.

• Runs locally on phones, single-board computers, and robots
• Text-focused, at about 230M parameters
• Aimed at low-latency, privacy-sensitive, and offline use cases

As always with a model this small, the practical question is how much capability survives the compression. Liquid AI's pitch is that a well-tuned tiny model can be good enough for a wide class of on-device tasks — and cheap enough to run almost anywhere. Full details are in the company's announcement.