Liquid AI opens LFM3: six 2B-class models off one 1.7B base

Liquid AI's pitch rests on a single backbone rather than a separate model per modality. The 1.7B base is a tri-hybrid of 13 convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers across 24 blocks, with a kernel size of 7 and a 128K vocabulary. That design choice is what lets one base spawn text, vision, audio and omni variants.

The company also positions LFM3 against its own earlier work. It says the 1.7B text model beats LFM2.5-2.6B while using 35% fewer parameters, posting 45.60 on LiveClaw, 71.39 on BFCL V4 tool-calling, 96.66 on IFEval, 80.36 on ToolSandbox and 52.83 on AA-LCR. Footprint is the other half of the argument: roughly 1 GB at 4-bit, and a single NVIDIA B300 can produce 3.6 billion tokens per day.

Malhar Palkar of Ambarella framed the constraint the release targets, saying the challenge is delivering these capabilities within the compute, memory and power limits of edge devices. The launch lands amid a steady run of nearly a hundred multimodal releases tracked over the past 90 days.