LFM3
Liquid AI released LFM3, a family of six 2B-class open-weight models for text, vision, audio, omni and decision tasks, built on one shared backbone and 22T tokens of pre-training.
LFM3 uses a new tri-hybrid architecture and brings all modalities onto one backbone, trained from LFM3-1.7B-Base. Liquid says LFM3-1.7B outperforms the earlier LFM2.5-2.6B with 35% fewer parameters. The text model has a 256K context window and fits in about 1 GB at 4-bit precision, and every model can be fine-tuned on a single GPU.
- Date
- Tuesday 27 October 2026
- Lab
- Liquid AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Pre-training tokens | 22T | company |
| Parameter reduction vs LFM2.5-2.6B | 35% fewer LFM3-1.7B said to outperform LFM2.5-2.6B | company |
| Text model context window | 256K | company |
| Text model memory at 4-bit | about 1 GB | company |
Base and post-trained models are on Hugging Face, the Liquid Playground and the Liquid AI API. License terms were not stated in the text read.
Sources
This record was checked against its sources on 28 October 2026. How we check