LFM3-1.7B
Liquid AI · 2026-10-27 · 1.7B parameters
LFM3-1.7B is Liquid AI's post-trained text model for local agents and specialized applications, released with open weights on October 27, 2026 as the text member of the first LFM3 family. It keeps LFM3-1.7B-Base's 24-layer tri-hybrid backbone of 13 gated short-convolution, five sliding-window attention and six global attention layers and its 262,144-token context, and was post-trained with SFT on 1.75T tokens, mixed reinforcement learning, multi-domain on-policy distillation from six specialist teachers, and agentic RL inside harnesses such as Hermes Agent, Pi and OpenClaw. Its chat template thinks by default and accepts enable_thinking=False, which the card warns can behave inconsistently. The BF16 checkpoint contains 1,728,801,536 parameters, and the tokenizer assigns 125,033 distinct IDs within 128,000 embedding rows.