Tether QVAC Releases Open-Source VisionPsy-Nano On-Device Multimodal Model

QVAC unveils 460M-parameter VisionPsy-Nano, achieving top normalized scores in on-device benchmarks. The Apache 2.0 licensed model enables private, efficient multimodal AI inference on edge hardware.

Woofun AI reports that QVAC, an AI research initiative under Tether Data, has open-sourced VisionPsy-Nano, a compact multimodal vision language model with approximately 460 million parameters designed for on-device and edge deployment. The model achieves the highest overall normalized score of 62.3 among evaluated on-device vision language models with fewer than 500 million parameters, outperforming competitors like Liquid AI and Hugging Face in 16 of 17 benchmark tests. Two versions are available under the Apache 2.0 license: the standard release and a Flash version optimized for low latency, which maintains around 99% of the original model's quality while delivering significantly faster inference speeds. Tether CEO Paolo Ardoino stated that this release demonstrates the viability of locally prioritized, efficient AI, allowing powerful and private systems to operate on existing user devices.