← All models

LFM2-VL-3B

Liquid AI · 2025-10-22 · 3B parameters

LFM2-VL-3B is Liquid AI's October 22, 2025 vision-language model built on the LFM2-2.6B language tower, whose 30 layers mix 22 gated short-convolution and eight grouped-query attention blocks, with a 400M-parameter SigLIP2 NaFlex vision encoder at native resolution. The card gives a 32,768-token text context and English support. The BF16 checkpoint contains 2,998,975,216 parameters, and the tokenizer assigns 64,400 distinct IDs within 65,536 embedding rows.

Benchmark scores