Liquid Foundation Models 3 Delivers Unified Multimodal Backbone for Edge Deployment

Liquid Foundation Models 3: A Unified Multimodal Backbone

Liquid AI has unveiled LFM3, the third generation of its Liquid Foundation Models, designed to serve as a single starting point for specialized intelligence across text, vision, audio, omni‑modal, and decision tasks. All six models sit in the 2‑billion‑parameter class and share a common 1.7‑billion‑parameter backbone, enabling developers to pick the modality they need, fine‑tune on proprietary data using a single GPU, and deploy wherever the application runs.

Key advances over LFM2

• Higher capability per parameter: a new tri‑hybrid architecture lets LFM3‑1.7B surpass the larger LFM2.5‑2.6B while using 35 % fewer parameters.
• All modalities on one backbone: six 2B‑class models—including the first omni model and video‑enabled vision—are derived from the same base.
• Edge‑ready footprint: the text model fits in roughly 1 GB at 4‑bit precision, custom vision and audio encoders accelerate inference, and a 256K‑token context window supports on‑device agents.
• Single‑GPU specialization: every variant can be fine‑tuned on one GPU, letting teams adapt terminology, data distributions, and deployment constraints without large clusters.

Why it matters for GPU and AI infrastructure

Because each LFM3 variant fine‑tunes on a single GPU, organizations can iterate quickly on domain‑specific models without provisioning multi‑node clusters. The compact 4‑bit footprint and long context window make it practical to run inference on edge GPUs or even NPUs, reducing latency and cloud egress costs. For AiGpu customers, this translates into lower compute spend per experiment and a clearer path from prototype to production on heterogeneous hardware.