3.1B total (LFM2.5-2.6B language backbone + 400M SigLIP2 NaFlex vision encoder)
32,768 tokens
LFM Open License v1.0
Open-weight vision-language model built through further mid- and post-training of LFM2-VL-3B, combining the LFM2.5-2.6B language backbone with a shape-optimized 400M-parameter SigLIP2 NaFlex vision encoder. The model processes text and one or more images at native resolution, splitting large images into non-overlapping 512x512 patches alongside a resized whole-image thumbnail, and supports 16 languages with a 128,000-token vocabulary. Liquid AI positions the release for direct, single-turn, high-throughput visual tasks, with improved natural-language object grounding, ScreenSpot-style GUI understanding, multi-image comprehension, and full-page OCR that emits experimental structured layout annotations with normalized bounding boxes, OTSL tables, and LaTeX equations. It also supports Pythonic structured function calls through dedicated tool-call tokens. Compared with LFM2-VL-3B, published results improve from 57.1 to 87.9 on RefCOCO, 50.2 to 61.5 on BLINK, 34.9 to 58.3 on MuirBench, 26.4 to 59.5 on ToolSandBox, 20.5 to 32.5 on BFCLv4, and 43.9 to 47.5 on the English subset of OCRBenchv2, while reaching 80.7 on ScreenSpot-v2. Liquid reports decoding speeds of 228 tokens per second on an Apple M5 Max, 116 tokens per second on an AMD Ryzen AI Max+ 395, and 20 tokens per second on a Galaxy S26 Ultra while fitting in about 3 GB of memory. On one H100 with vLLM, the model reportedly reaches approximately 11,000 output tokens per second at high concurrency and about 34 milliseconds time to first token for a five-frame clip. Official native, GGUF, ONNX, and MLX releases support Transformers, vLLM, SGLang, llama.cpp, WebGPU, and Apple Silicon deployment. Liquid does not recommend the model for long-context or reasoning-intensive visual tasks.