LFM2.5-VL-3B
Liquid AI · 2026-08-12 · 3.1B parameters
LFM2.5-VL-3B is Liquid AI's August 12, 2026 vision-language model, built on LFM2-VL-3B with further mid- and post-training on the LFM2.5-2.6B backbone and a 400M-parameter shape-optimized SigLIP2 NaFlex vision encoder; it takes text and one or more images and is a non-reasoning model. The card and configuration give a 32,768-token context and 16 languages. The BF16 checkpoint contains 3,123,483,888 parameters, and the tokenizer assigns 125,017 distinct IDs within 128,000 embedding rows.