LFM3-1.7B
★ agentic
Overview
Liquid's tri-hybrid LFM3 architecture (gated convolutions + sliding-window + global attention) post-trained for local agents - agent scores that rival 9B models, a 256K context window, and 2.5x faster prefill than same-size peers on edge hardware. Official BF16 checkpoint for vLLM and SGLang.
Good for
- agentic
Recommended hardware
≈ 8 GB VRAM to run this 1.7B model at the BF16 quantization.
Fits 8 GB+ consumer GPUs (e.g. RTX 4060, GTX 1070), or CPU inference