Liquid releases LFM3, a family of small models for devices

The makers of Liquid Foundation Models released LFM3, a new family of small AI models for text, vision, audio and more. All are in the 2B parameter range. A parameter is a number that a model learns. You can train each model more on your data with one graphics processor. The text model fits in about 1 GB at 4-bit precision.

What is new

All models start from one backbone, LFM3-1.7B-Base. The text model has a context window of 256K tokens, and agents can operate on a device. The family has:

• a text model
• a vision model that takes video
• an audio model for voice in 18 languages
• an omni model that sees, hears and talks

The numbers

LFM3-1.7B has 35% fewer parameters than LFM2.5-2.6B. The vision model raises its average score from 57.6 to 71.4. The makers say it matches Gemma 4 12B, a model more than six times larger.

What is not known

All numbers come from the makers. The source shows no test by an external group. It gives no result for your tasks or your devices.

Insight: Small models on devices and giant open models go in opposite directions

On 27 October, Ant Group put a 574 GB open model online. The text model of Liquid fits in about 1 GB. Maybe the market has two paths. One path sells large models. The other path sells models that you adapt to your data. A team can test the small path at low cost.