Liquid AI's LFM3: six 2B-class models on one backbone

LFM3 follows Liquid AI's earlier LFM2 family, which showed that a hybrid of short convolutions and attention can run fast on-device without sacrificing quality. For LFM3, the company redesigned that architecture through a hardware-in-the-loop search now covering 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. The result is a base model, LFM3-1.7B-Base, from which all six models branch.

The efficiency gains are the through-line. LFM3-1.7B is said to outperform LFM2.5-2.6B while using 35% fewer parameters, with the biggest gains on agentic tasks: 38.42 against 3.02 on WebArena-Verified. The vision model raises its average across 62 vision benchmarks from 57.6 to 71.4 compared with LFM2.5-VL-3B, and the audio model improves speech quality on UTMOSv2 to 3.83 from 3.22. Two new in-house encoders underpin the multimodal models, including LiquidCodec, an audio tokenizer that turns generated tokens into speech at about 87x real time on a single CPU thread on Apple M4 Max.

The release also points beyond benchmarks to deployment. Liquid AI adapted LFM3-VL-2B into three robotics models for orchestration, navigation and manipulation; the manipulation model achieves 96.2% average success across the four LIBERO suites and generates action chunks in under 50 ms on a Jetson AGX Orin with TensorRT FP16. The pitch is that a short fine-tune can turn a small model into a reliable expert on a specific task, deployed from devices to private infrastructure to the cloud.