Liquid AI's LFM3 puts six models on one 2B backbone

A tri-hybrid architecture lets a single 1.7B base model branch into text, vision, audio, omni, and decision variants—all fine-tunable on one GPU.

By The ai-compute Desk

Liquid AI has released LFM3, a family of six models that share a single 1.7B-parameter backbone and cover text, vision, audio, omni, and decision tasks—all in the 2B-parameter range. [1]

The architecture is a tri-hybrid stack: 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. [1] That design lets LFM3-1.7B outperform the previous LFM2.5-2.6B while using 35% fewer parameters, with the sharpest gains on agentic tasks—scoring 38.42 against 3.02 on WebArena-Verified. [1]

Edge deployment is a first-class constraint. The text model fits in about 1 GB at 4-bit precision, and a 256K context window supports on-device agents. [1] A new audio encoder keeps time to first token as low as 50 ms, while LiquidCodec generates speech at about 87x real time on a single CPU thread on Apple M4 Max. [1]

The vision model, LFM3-VL-2B, raises the average across 62 vision benchmarks from 57.6 to 71.4 compared with its predecessor—on par with Gemma 4 12B, a model more than six times larger. [1] A robotics fine-tune of the same base achieves 96.2% average success across the four LIBERO suites, generating action chunks in under 50 ms on a Jetson AGX Orin. [1]

The bet is that modality-specific small models, trained from one shared base, beat a single large generalist when your task is narrow and your hardware is constrained.

All LFM3 models are open-weight and available on Hugging Face now, with fine-tuning notebooks covering text, vision, audio, and omni tasks. [1]

Frontier models are powerful generalists, but they are not necessarily experts in your use case.Liquid AI