Tether Unveils Open-Source On-Chain Visual Language Model VisionPsy-Nano

July 29th, according to the Tether blog, Tether's AI research project QVAC under Tether Data has open-sourced VisionPsy-Nano. This is a compact visual language model designed for on-device and edge deployment, with a parameter size of about 460 million. It can directly bring the previously cloud-reliant multimodal understanding capability to mobile devices. Among all on-device visual language models with less than 500 million parameters that have been evaluated, this model achieved the highest overall normalized score of 62.3 points. In 16 out of 17 benchmark tests, it outperformed competing models such as Liquid AI and Hugging Face.

Tether released two versions this time, with the Flash version optimized for low latency, greatly improving the inference speed while retaining about 99% of the full model quality. Both versions are open-sourced under the Apache 2.0 license. Paolo Ardoino, CEO of Tether, stated that this proves that a local-first, efficient AI is a feasible path, allowing powerful and private AI to run on people's existing devices.