Liquid AI opens six-model LFM3 family at 256K context

Liquid AI released LFM3 on October 27, a family of six 2B-class open-weight models covering text, vision, audio, omni and two decision variants, all branched from one 1.7B base. In its launch post, the company writes that "the first LFM3 release includes six models, all branched from the same LFM3-1.7B-Base backbone," and that "LFM3-1.7B-Base was pre-trained on 22T tokens." The text model carries a 256K context window.

The backbone is a tri-hybrid of gated short-convolution layers, sliding-window attention and global attention. The mix is 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. Liquid says the 1.7B text model beats its own prior 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.

Where the pitch lands is footprint. The 1.7B base fits in roughly 1 GB at 4-bit precision. The post reports that "LFM3-1.7B is the fastest model on all three devices, decoding 1.8 to 2.8 times faster than the next one," and that a single NVIDIA B300 can produce 3.6 billion tokens per day. Supported targets listed include Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI and Qualcomm Hexagon NPU.

Ambarella's Senior Director for AI, Malhar Palkar, framed the brief in a launch comment: "The challenge is delivering these capabilities within the compute, memory, and power constraints of edge devices." "The base and post-trained models are open-weight and available today on Hugging Face," the post states. The release lands amid a steady run of nearly a hundred multimodal releases we have tracked over the past 90 days.