Liquid AI's LFM2-VL Runs Vision AI on a Phone at 450M Parameters
Liquid AI's LFM2-VL-450M brings multimodal vision to edge devices with 2x GPU inference speed and under 500M parameters
Takeaways
- LFM2-VL-450M released: Liquid AI launches its first vision-language model series, with a 450M and 1.6B variant targeting edge hardware. [HF]
- 2x faster inference: Both models deliver up to 2x GPU inference speed compared to comparable VLMs like SmolVLM2-500M.
- Edge-first design: Models run on phones, wearables, and embedded systems; a WebGPU demo runs entirely in-browser.
- Strong instruction following: LFM2-VL-450M scores 33.09 on MM-IFEval vs. 11.27 for SmolVLM2-500M at similar parameter count.
- Inference-time speed/quality tradeoff: Users can tune image token count at runtime without retraining the model.
- Free for small companies: Apache 2.0-based license allows commercial use for companies under $10M revenue; larger companies need a paid license.
Liquid AI has released LFM2-VL, its first series of vision-language models, and the headline number is striking: a fully capable image-and-text model in just 450 million parameters. That's small enough to run on a phone, a wearable, or an embedded system -- and it still beats comparable tiny VLMs on most benchmarks.
LFM2-VL is Liquid AI's first series of multimodal foundation models, designed for low-latency and device-aware deployment. It comes in two variants: the hyper-efficient LFM2-VL-450M for highly resource-constrained settings, and the more capable but still lightweight LFM2-VL-1.6B. Both weights are available now on Hugging Face.
Why edge VLMs matter right now
Most large VLMs require large GPU memory and cloud infrastructure to run -- a real problem for deployment scenarios like warehouse robots, smart glasses, or retail shelf cameras, where computing is limited and latency is critical. The race to shrink capable vision models to fit this hardware is intensifying, and LFM2-VL is Liquid AI's entry into that race.
LFM2-VL is designed for efficient deployment across a wide range of hardware -- from smartphones and laptops to wearables and embedded systems -- promising low-latency performance, strong accuracy, and flexibility for real-world applications.
The architecture doing the heavy lifting
LFM2-VL consists of three main components: a language model backbone, a vision encoder, and a multimodal projector. Each component was chosen specifically to minimize compute while preserving quality.
- Language backbone: LFM2-VL-450M builds on the LFM2-350M backbone, while LFM2-VL-1.6B uses LFM2-1.2B.