Liquid AI has released LFM2-VL-450M, a 450-million-parameter vision-language model designed for on-device inference. Built on Liquid’s LFM2-350M language backbone and an 86M-parameter SigLIP2 NaFlex vision encoder, the model delivers 2× faster inference than comparable VLMs on GPUs while processing images at native resolution — all under an Apache 2.0 license.
Intermediate
Key points
- Liquid AI released LFM2-VL-450M, a 450-million-parameter vision-language model for on-device inference that delivers 2× faster inference than comparable VLMs on GPUs under an Apache 2.0 license.
- It combines an LFM2-350M language tower, an 86M-parameter SigLIP2 NaFlex vision encoder, and a 2-layer MLP multimodal projector with pixel unshuffle for token reduction.
- The vision encoder processes images at native resolution up to 512×512 pixels without upscaling, handles non-standard aspect ratios without distortion, and splits larger images into non-overlapping 512×512 patches.
- A 256×384 image generates approximately 96 visual tokens, keeping the context budget lean for on-device use.
- Liquid AI says LFM2-VL-450M posts OCRBench 655, SEEDBench_IMG 63.5, and RealWorldQA 52.29 despite its small parameter count.
Architecture
LFM2-VL-450M combines three components:
- Language tower: LFM2-350M, Liquid AI’s compact language backbone based on their proprietary Liquid Foundation Model architecture
- Vision tower: SigLIP2 NaFlex encoder (86M parameters, base variant) for fast image processing
- Multimodal projector: A 2-layer MLP with pixel unshuffle for token reduction
The vision encoder processes images at their native resolution up to 512×512 pixels without upscaling, and handles non-standard aspect ratios without distortion. Larger images are split into non-overlapping 512×512 patches. A 256×384 image generates approximately 96 visual tokens — keeping the context budget lean for on-device use.
Performance
Despite its tiny parameter count, LFM2-VL-450M posts competitive benchmark numbers:
- OCRBench: 655 — strong document understanding for a sub-500M model
- SEEDBench_IMG: 63.5 — solid image understanding
- RealWorldQA: 52.29 — practical visual reasoning
- Inference speed: 2× faster than comparable VLMs on GPU
The model features user-tunable speed-quality tradeoffs at inference time, allowing developers to optimize for their specific latency and accuracy requirements. It was trained on approximately 100 billion multimodal tokens.
What This Means
At 450 million parameters, LFM2-VL sits in a class of models designed to run directly on smartphones and edge devices. While larger VLMs like Qwen-VL and LLaVA dominate benchmarks, they require server-grade hardware. Liquid AI’s bet is that many vision-language tasks — OCR, document scanning, visual QA — don’t need a 7B+ model. The Apache 2.0 license and GGUF availability on Hugging Face make it easy to integrate into mobile and embedded applications.
This post was drafted with AI assistance and reviewed by RITS staff.
Leave a Reply