Tether Data introduced VisionPsy-Nano, a compact vision-language AI model with 460 million parameters designed to run directly on consumer devices like smartphones and edge gadgets. Developed by Tether's QVAC AI research team, this model processes images and text entirely on-device, cutting out the need for cloud servers.
The release offers two variants. VisionPsy-Nano-460M focuses on delivering the best quality at its parameter size. The Flash version is optimized for speed, enhancing mobile responsiveness while maintaining nearly all original capabilities.
Performance that Challenges Cloud-Based AI
On flagship Android phones, the Flash model delivers first-token generation speeds 19 to 23 times faster than comparable models like SmolVLM2-500M. On the iPhone 15, speed gains soared to 36 times. Tether CEO Paolo Ardoino said this proves that powerful AI doesn’t have to rely on centralized data centers anymore.
VisionPsy-Nano scored highest among models under 500 million parameters with a normalized score of 62.3. It outperformed competitors from Liquid AI and Hugging Face on 16 out of 17 benchmark tests. The model particularly excelled in document understanding, OCR, visual perception, reasoning, and reliability, with a 4.6% lead in visual tasks and a 7.4% edge in reasoning over others in its class.
By enabling developers to integrate AI directly on user devices, Tether aims to bypass centralized AI providers and enhance privacy for end-users. This strategy aligns with broader trends favoring on-device AI for faster, more secure applications.
This content is for informational purposes and not investment advice.
