← All models

LFM3-1.7B-Base

Liquid AI · 2026-10-27 · 1.7B parameters

LFM3-1.7B-Base is the pre-trained text-only checkpoint behind LFM3-1.7B and the rest of Liquid AI's first LFM3 family, released with open weights on October 27, 2026 and recommended for heavy fine-tuning or new post-training. Its 24-layer tri-hybrid backbone has 13 gated short-convolution layers with kernel size 7, five sliding-window attention layers with a 256-token window and six global attention layers, with sigmoid-gated attention heads, partial RoPE on the global layers and sandwich normalization. Liquid reports 22T pre-training tokens in total, the last 3.5T a mid-training stage that extends context to 262,144 tokens, and its chat template thinks by default unless enable_thinking=False. The BF16 checkpoint contains 1,728,801,536 parameters, and the tokenizer assigns 125,033 distinct IDs within 128,000 embedding rows.

Benchmark scores

No benchmark scores recorded.