Liquid AI drops fast, tiny vision models for edge devices

LFM2-VL is Liquid AI's first lineup of open-weight vision-language models, bringing its Liquid Foundation Model architecture directly to multimodal edge computing. Available in 450M and 1.6B parameter sizes, it is built to run on phones, laptops, wearables, and embedded hardware with text and variable-resolution image inputs.

How it works under the hood

LFM2-VL consists of a language backbone, a vision encoder, and a multimodal projector.

The language towers are inherited from the LFM2 family, using LFM2-350M for the 450M variant and LFM2-1.2B for the 1.6B variant.

For vision, Liquid AI uses SigLIP2 NaFlex encoders. The 450M model uses a fast 86M base encoder, while the 1.6B model integrates a 400M shape-optimized encoder for fine-grained image understanding. The system processes native inputs up to 512x512 pixels without forced upscaling or aspect-ratio distortion.

Larger images are sliced into non-overlapping 512x512 patches. The 1.6B version also feeds the model a low-resolution thumbnail of the whole image to preserve global scene context. Special tokens track patch coordinates and mark the thumbnail boundaries.

The projector combines a 2-layer MLP with pixel unshuffle to keep vision token counts low. A 256x384 image translates into just 96 tokens, a 384x680 image takes 240 tokens, and a 1000x3000 image uses 1,020 tokens.

Critically, you can tune the maximum number of vision tokens and patch counts at inference time without retraining. That means you can dial in your own latency and accuracy tradeoffs based on target hardware.

Both models were trained on roughly 100 billion multimodal tokens, shifting the text-to-image ratio from 95% down to 30% during mid-training.

Speed and benchmark results

Throughput is the big story here. In a benchmark processing a 1024x1024 image with a short prompt and generating 100 output tokens, LFM2-VL ran up to 2x faster on GPU than competing small VLMs.

On vision benchmarks, LFM2-VL-1.6B scored 65.23 on RealWorldQA, matching InternVL3-2B and beating SmolVLM2-2.2B at 57.50. It scored 37.66 on MM-IFEval and 742 on OCRBench.

The tiny LFM2-VL-450M scored 52.29 on RealWorldQA and 655 on OCRBench, outperforming the similar-sized SmolVLM2-500M across most core metrics.

How to get started

Model weights are live on Hugging Face alongside a Colab fine-tuning notebook.

LFM2-VL works with Hugging Face transformers and TRL out of the box. Enterprise teams can also deploy and tune models across hardware via Liquid AI's LEAP platform.

The release uses an Apache 2.0-based open license. Academic and research use is free, and commercial use is entirely free for businesses generating under $10M in annual revenue. Larger enterprises must contact sales for licensing.

Why it matters

If you build on-device assistants, smart glass apps, or offline robotics, LFM2-VL is worth immediate attention. Deploying vision models on edge hardware usually forces brutal compromises on RAM or latency, but sub-2B footprints make local deployment genuinely practical.

We think Liquid AI made a sharp bet here. Instead of chasing bloated multimodal leaderboards, it attacked latency bottlenecks with smart token compression and flexible patch handling. It won't out-reason a cloud frontier model on complex math, but for edge applications where every millisecond counts, up to 2x speedups are impossible to ignore.