Liquid AI has released LFM2-VL-3B, a 3 billion parameter vision-language model designed to run on edge hardware rather than in a data center. It is the largest model in the company’s LFM2-VL family, which previously topped out at 1.6B parameters, and is available now under the LFM Open License on Hugging Face and through Liquid AI’s LEAP platform.
The model pairs Liquid AI’s densest text backbone, LFM2-2.6B, with a SigLIP2 400M NaFlex vision encoder. That combination lets it take in images at their native resolution and aspect ratio, then compress them to a configurable number of vision tokens. Developers can trade quality for speed by adjusting how many tokens each image consumes, a knob that matters when the model has to fit on a phone, a workstation, or an on-premise server.
On a standard suite of open multimodal benchmarks, LFM2-VL-3B edges out larger and similarly sized peers. It scores 67.31 on the aggregate the company reports, ahead of InternVL3.5-2B at 66.63, Qwen2.5-VL-3B at 66.61, and InternVL3-2B at 66.46. It reaches 71.37 on RealWorldQA and 51.83 on MM-IFEval, with low hallucination on the POPE benchmark and strong English OCR. Liquid AI notes it left Qwen3-VL-2B out of the table because that model shipped the day before.
Language knowledge holds up relative to the text backbone, at 30% on GPQA and 63% on MMLU. More notably, the model reads images in nine languages beyond English: Japanese, French, Spanish, German, Italian, Portuguese, Arabic, Chinese, and Korean. That positions it for documents and interfaces that mix text and visuals across locales.
The release fits a broader shift toward small, efficient multimodal models that can run where the data lives. A 3B model that handles images on device avoids sending sensitive scans or documents to the cloud, a constraint that matters in healthcare, manufacturing, and other regulated settings. Liquid AI frames the LFM2 line as proof that architecture and training choices, not just scale, still move the needle on practical deployments.
LFM2-VL-3B joins 450M and 1.6B variants in the family, giving developers a range of sizes for different hardware. All are tuned for instruction following and lightweight agentic flows rather than safety-critical decisions.
Sources
Liquid AI, “LFM2-VL-3B: A New Efficient Vision-Language for the Edge” (Aug 2026): https://www.liquid.ai/blog/lfm2-vl-3b-a-new-efficient-vision-language-for-the-edge
LFM2-VL-3B model card, Hugging Face: https://huggingface.co/LiquidAI/LFM2-VL-3B
