Liquid AI releases LFM2-VL vision-language models
Liquid AI has released LFM2-VL, a pair of open-weight 450M- and 1.6B-parameter vision-language models built for phones, laptops and embedded devices, claiming up to 2× faster GPU inference than comparable models.
Liquid AI announced LFM2-VL, its first vision-language series and an extension of its open-weight LFM2 family. The models take text and images at variable resolutions and target low-latency deployment on devices from single GPUs down to wearables.
There are two variants. LFM2-VL-450M pairs the LFM2-350M language backbone with an 86M SigLIP2 NaFlex base encoder. LFM2-VL-1.6B pairs LFM2-1.2B with a 400M shape-optimized SigLIP2 encoder. Images are processed at native resolution up to 512×512. Larger images are split into non-overlapping 512×512 patches, and the 1.6B model also receives a downscaled thumbnail for global context. A 2-layer MLP projector with pixel unshuffle cuts the image token count. For example, a 1000×3000 image yields 1,020 tokens. Users can cap image tokens and patch counts at inference time to trade quality for speed without retraining.
Training ran on roughly 100 billion multimodal tokens. During joint mid-training, the share of text data was stepped down from 95% to 30%, and supervised fine-tuning followed.
On Liquid's own benchmarks, the 1.6B model scores 65.23 on RealWorldQA, slightly ahead of InternVL3-2B's 65.10. It trails InternVL3-2B on most other tests, including MMStar (49.53 vs 61.10), OCRBench (742 vs 831) and MMMU (38.44 vs 48.70). Liquid says LFM2-VL runs up to 2× faster than the fastest comparable model on a workload of one 1024×1024 image and 100 output tokens. It did not publish the speed figures for each model.
The weights are on Hugging Face under an Apache 2.0-based license. Commercial use is free for companies with under $10M in revenue. Larger companies must obtain a commercial license from Liquid.