Tether Data Unveils VisionPsy-Nano: A Breakthrough in On-Device AI

Tether Data has made significant strides in artificial intelligence (AI) research by open-sourcing its vision-language model, VisionPsy-Nano. This 460-million-parameter model is designed to run directly on smartphones, eliminating the need for cloud servers and allowing for offline processing of images, documents, and charts.

VisionPsy-Nano outperformed competitors across 16 of 17 benchmark tests, including Liquid AI's LFM2.5-VL-450M and Hugging Face's SmolVLM2-500M. The model's 'Flash' version ran up to 36 times faster than SmolVLM2-500M on an iPhone 15.

The VisionPsy-Nano model supports document analysis, optical character recognition, and extraction of text and structure from financial reports, flowcharts, and infographics. QVAC says the model retained about 99% of its performance while producing faster responses.

Tether's AI research arm, QVAC, has published its benchmark configurations through VLMEvalKit, allowing outside researchers to repeat the tests. The model is available for commercial use under the Apache 2.0 license.