Liquid AI releases LFM3, six open-weight ~2B models spanning text, vision, audio, omni and decision tasks

Background and pain point

Small on-device models usually lack the capability for agent work, while cloud models bring latency, cost and privacy trade-offs. Liquid AI's LFM2 showed a convolution-plus-attention hybrid can run fast on phones; LFM3 aims to pack more agentic capability into the same size class.

Architecture

Liquid redesigned the network with hardware-in-the-loop search across 14 device and backend combinations, including the Samsung Galaxy S26 Ultra, Raspberry Pi 5, Apple M4 Max and the Qualcomm Hexagon NPU. Of 24 layers, 13 are gated short convolutions (kernel size raised from 3 to 7), 5 use sliding-window attention with a 256-token window and 6 use global attention, plus head-wise gating, partial RoPE and sandwich norm. Pre-training used 22T tokens, with a final 3.5T-token mid-training stage rich in agent traces that also extends context to 256K.

Benchmarks

All figures are vendor-reported: LFM3-1.7B scores 71.39 on BFCL V4, 96.66 on IFEval and 38.42 on WebArena-Verified, beating Qwen3.5-9B on the latter agent tasks. It still trails larger models on SWE-bench Verified (43.40 vs 51.20) and GPQA Diamond (63.37 vs 80.40), and Liquid itself says size matters most for coding and knowledge. No independent third-party evaluation is available yet.

Getting started

Weights are in the LFM3 collection on Hugging Face with GGUF, MLX, ONNX, FP8 and NVFP4 variants, and the models are also on the Liquid Playground and API. Liquid says every model can be fine-tuned on a single GPU, making them candidates for specialized agents such as inbox triage, calendars or tool search. The license is the LFM Open License v1.0, not MIT or Apache, so check the terms before commercial use.