LFM3-VL-2B

LFM3-VL-2B is Liquid AI's open-weight 1.92B vision-language model (1.73B LFM3 language model + new 164M in-house conv/attention vision encoder + 25M projector), released Oct 27, 2026 as part of the LFM3 family. Inputs: text, images, and video (first LFM vision model with video, sampled at 2 fps) at native resolution; text output. Reasons step by step before answering (thinking), with spatial/embodied reasoning and grounding aimed at robotics, plus document/OCR/chart and GUI understanding. 128K context; 16 languages. Vendor reports a 71.4 average across 62 vision benchmarks (vs 57.6 for LFM2.5-VL-3B). Not recommended for coding, multi-turn conversations, or complex reasoning per the model card. Hugging Face: LiquidAI/LFM3-VL-2B (+ GGUF, MLX, ONNX). License: LFM Open License v1.0. Catalog benches are the thinking-mode column from the model card table. No catalog id (recorded here only): SEED-Bench image 78.6, MM-Vet v1 57.4, PerceptionBench 37.1 (catalog id is Moonshot's internal benchmark), HR-Bench 8K 76.2, ZoomBench 62.6, V*Bench 86.4, CV-Bench 88.8, LogicVista 60.4, RefCOCO-M 95.2, Point-Bench 73.8, MindCube tiny 56.5, SAT 80.2, Where2Place 44.9, OpenEQA 76.3, VSI-Bench 63.1, ScreenSpot-v2 desktop/mobile/web 94.6/92.2/89.9, MMMB 85.2, Multilingual MMBench 83.6, MM-ToolSandbox 42.5. No gateway id or price at time of adding, so no provider row. Sources: https://huggingface.co/LiquidAI/LFM3-VL-2B · https://www.liquid.ai/blog/lfm3.