Tether Data launches VisionPsy‑Nano, a 460‑M parameter AI model for on‑device vision tasks

Tether Data’s AI research unit QVAC released VisionPsy‑Nano, an open‑weight 460‑million‑parameter vision‑language model built to run directly on smartphones and other edge devices. The model is offered under the Apache 2.0 license and comes in two variants: the standard VisionPsy‑Nano‑460M, which prioritises accuracy, and VisionPsy‑Nano‑460M‑Flash, which sacrifices a small amount of quality for speed, delivering up to 36 times faster first‑token generation on an iPhone 15 compared with competing models.

In benchmark testing the model achieved the highest normalized score (62.3) among sub‑0.5 B parameter systems, outperforming rivals from Liquid AI and Hugging Face on 16 of 17 tests, including document understanding, OCR, visual reasoning and reliability tasks. CEO Paolo Ardoino said the release shows that “local‑first, highly efficient AI is a viable pathway,” highlighting the move away from cloud‑centric AI toward privacy‑preserving, offline processing on consumer devices.