Tether AI Research introduced VisionPsy-Nano, a family of compact roughly 460-million-parameter vision-language models designed for on-device deployment on mobile hardware. The two variants, a quality-focused model and a latency-optimized ‘Flash’ variant, achieve state-of-the-art results for their size and outperform larger competing models on benchmarks such as instruction following and reasoning. The Flash variant reaches time-to-first-token roughly 19-23x faster than comparable models on smartphones through reduced visual token processing, while retaining about 99% of the full model’s accuracy.
