Liquid AI’s LFM2.5-230M Shakes Up On-Device Language Models

Liquid AI just dropped a bombshell in the world of language models. Their newest creation, the LFM2.5-230M, is the company’s smallest and fastest yet. This model isn’t just small—it’s powerful enough to beat much larger rivals on key benchmarks. And it runs directly on devices you already own, from flagship smartphones to tiny Raspberry Pis. What does that mean? On-device AI that’s fast, efficient, and ready for real-world use.

Speed and Size That Defy Expectations

The LFM2.5-230M comes packing 230 million parameters, built on Liquid AI’s LFM2 architecture. That might sound modest compared to models with billions of parameters. But don’t let size fool you. This tiny titan blazes through tasks at 213 tokens per second on a Galaxy S25 Ultra. Even on a Raspberry Pi 5, it keeps pace at 42 tokens per second. That’s a huge deal for edge devices where power and speed are precious.

Behind this speed is a massive training effort. The model learned from 19 trillion tokens during pre-training, including a phase that extended its context window to 32,000 tokens. That means it understands much longer text snippets, perfect for data extraction and instruction following.

Multilingual Mastery and Benchmark Domination

Liquid AI designed the LFM2.5-230M to speak ten languages, including English, Chinese, Arabic, and Japanese. This makes it a versatile tool for global applications. But it’s not just about languages. The model excels on tough benchmarks:

• IFEval: 71.71
• IFBench: 38.40
• CaseReportBench: 22.51

These scores put it ahead of much larger competitors like Qwen3.5-0.8B, which scored 59.94 on IFEval, and Gemma 3 1B IT with 63.49. That’s right—this smaller model outperforms models four to five times its size in instruction following and data extraction tasks.

Liquid AI is clear about the model’s strengths and limits. They say, “The pitch is narrow on purpose. This is not a general reasoning model. It is built for data extraction and tool use on edge hardware.” So, if you want to crunch advanced math, generate code, or write creatively, this isn’t the model for you.

Open Access Meets Commercial Boundaries

Liquid AI released the LFM2.5-230M on Hugging Face, supporting multiple popular runtimes like llama.cpp, MLX, vLLM, SGLang, and ONNX. Developers and researchers can tap into this model’s power right away, benefiting from on-device inference capabilities.

But there’s a catch for big players. The model operates under the LFM Open License v1.0, a restricted dual-use commercial license. Companies making $10 million or more annually must negotiate separate commercial agreements to use it. This strikes a balance between openness and control, ensuring responsible use.

Real-World Impact and What’s Next

One exciting application of LFM2.5-230M is its deployment on the Unitree G1 humanoid robot. This shows the model’s ability to power robotics with efficient language understanding on the edge. Plus, the release follows close behind Liquid AI’s 350M multilingual search models introduced just days earlier.

Liquid AI’s approach prioritizes lean, specialized AI that runs where users need it most—right on their devices. This shift to efficient edge AI is game-changing. It means less reliance on cloud servers, faster responses, and better privacy.

The future looks bright for compact, task-focused models. As AI pushes into every device, Liquid AI’s LFM2.5-230M sets a new standard for what small can do. Who says bigger is always better?