Liquid AI releases LFM2-VL-3B vision-language model
Liquid AI's new 3-billion-parameter vision-language model aims for higher accuracy on edge devices while keeping its LFM2 architecture fast, and the company's own benchmarks put it narrowly ahead of similar-sized open models.
Liquid AI has released LFM2-VL-3B, the largest model so far in its vision-language family. The earlier models have 450M and 1.6B parameters. The new model is available on Hugging Face under the company's LFM Open License and through its LEAP platform.
The model uses the same recipe as Liquid's earlier VLMs. It pairs the company's LFM2-2.6B dense language backbone with a SigLIP2 400M NaFlex vision encoder. That setup lets it handle images at their native resolution and aspect ratio. Developers can also change the number of vision tokens per image to choose between accuracy and speed, which matters most when deploying on edge hardware.
In Liquid's own evaluations, LFM2-VL-3B averaged 67.31 across 13 benchmarks. That puts it just ahead of InternVL3.5-2B (66.63), Qwen2.5-VL-3B (66.61) and InternVL3-2B (66.46), and well ahead of SmolVLM2-2.2B (54.85). Its clearest leads were on RealWorldQA (71.4% vs. 65.2% for Qwen2.5-VL-3B) and MM-IFEval (51.8% vs. 47.3% for InternVL3.5-2B). It scored lower than the comparison models on MMMU (45.33) and MME.
Liquid says the model keeps about the same text-only knowledge as its backbone, with 30% on GPQA and 63% on MMLU. It also says the model has low hallucination rates on POPE. Visual understanding now covers Japanese, French, Spanish, German, Italian, Portuguese, Arabic, Chinese and Korean as well as English.
Liquid did not publish latency or on-device throughput figures. No independent evaluations have been published yet.