Liquid AI LFM3: six small open models that run text, vision and voice on your own device

Most AI headlines this month have been about giant frontier models, so I was happy to see a release going the other way. Liquid AI has launched LFM3, a family of open-weight models in the 2B parameter range that are built to run on phones, laptops and small boards rather than in a data center.

Six small models, one backbone

All six models branch from the same base model, LFM3-1.7B-Base, which Liquid AI says was trained on 22 trillion tokens:

• LFM3-1.7B for text, with a 256K-token context window and a footprint of about 1 GB at 4-bit precision
• LFM3-VL-2B for images and, for the first time in this line, video
• LFM3-Audio-2B for voice chat, transcription and speech in 18 languages
• LFM3-Omni-2B, which can see, listen and talk back in one model
• d1-2B and d1-Omni-2B, decision models that answer a list of questions about an input in a single pass

The company says the text model beats its own earlier 2.6B model while using 35% fewer parameters, and that every model in the family can be fine-tuned on a single GPU. The weights are on Hugging Face, with day-one support in llama.cpp, MLX and vLLM.

Why I care

For anyone who games or creates on a laptop, small models like these are what make AI features possible offline, without sending your screen, voice or files to a server. Liquid AI tested on hardware from Samsung and OnePlus phones to a Raspberry Pi 5, Apple and AMD laptops, and Qualcomm's NPU, and lists AMD, Qualcomm, NVIDIA, Arm and Intel among its launch partners.

These are the company's own benchmarks, so I would wait for independent testing before reading too much into the charts. Still, a capable voice and vision assistant that fits in a couple of gigabytes is a real step for local AI.

Source: Liquid AI blog