Liquid AI Releases LFM3, Six Open-Weight Models Claiming a 12.7x Agentic Benchmark Jump on a 1.7B Backbone
The tri-hybrid architecture pairs gated short convolutions with sliding-window and full attention, cutting KV cache to about 27% of an equivalent transformer.
Liquid AI released LFM3 on 27 October, a family of six open-weight multimodal models built on a 1.7-billion-parameter backbone pre-trained on 22 trillion tokens with a 256K context, under the LFM Open License v1.0. The architecture is a tri-hybrid: 13 gated short convolution layers with O(1) state and no KV cache, 5 sliding-window attention layers with a 256-token window, and only 6 full attention layers, cutting KV cache to roughly 27% of an equivalent 24-layer transformer. The final 3.5 trillion training tokens were mid-training on agentic traces and tool calls.
The company reports LFM3-1.7B scoring 38.42 versus 3.02 for LFM2.5-2.6B on WebArena-Verified, a 12.7x jump on agentic web tasks. The 4-bit text model is about 1.1GB and fine-tunable on a single consumer GPU, targeting edge devices and small servers. All benchmark figures are self-reported and have not yet been independently reproduced.