← All models

LFM3-VL-2B

Liquid AI · 2026-10-27 · 1.9B parameters

LFM3-VL-2B is Liquid AI's post-trained vision-language model in the first LFM3 family, released with open weights on October 27, 2026 and the first LFM vision model to take video, sampled at 2 fps with 4 to 768 frames by default. It pairs the 1,728,801,536-parameter LFM3 language model with a new 164,206,336-parameter in-house vision encoder, 14 layers of gated depthwise 2D convolutions and bidirectional attention over 16×16 patches, and a 25,174,016-parameter projector, for 1,918,181,888 BF16 parameters. It thinks before answering by default, and its card also reports a configuration with thinking disabled. The card and configuration give a 128,000-token context, and the tokenizer assigns 125,031 distinct IDs within 128,000 embedding rows.

Benchmark scores