Tether Data has launched VisionPsy-Nano, an open-weight 460 million-parameter vision-language model developed by its QVAC AI research initiative and optimized to run directly on smartphones and edge devices.

VisionPsy-Nano is designed to deliver multimodal AI without relying on cloud infrastructure, according to Tether. The model allows devices to process images, text and visual information locally.

The release includes two versions built for different use cases. As noted by Tether, VisionPsy-Nano-460M delivers the highest possible quality at its size.

VisionPsy-Nano-460M-Flash, meanwhile, is optimized for speed, preserving nearly all of the original model’s performance while improving mobile responsiveness.

The Flash model targets everyday device deployment, delivering first-token generation improvements of roughly 19 to 23 times compared with SmolVLM2-500M and Tether’s base model on several Android flagship devices. Performance gains reached up to 36 times on iPhone 15.

Tether CEO Paolo Ardoino said the release demonstrates that efficient local AI can challenge the industry’s reliance on centralized data centres.

“Achieving best-in-class quality and performance on general vision tasks at just 460 million parameters proves that local-first, highly efficient AI is a viable pathway. By putting these tools directly into the hands of developers, we are bypassing centralized gatekeepers and making powerful, private AI possible on the devices people already own,” Ardoino stated.

The company said VisionPsy-Nano sets a new benchmark for compact vision-language AI, ranking first among evaluated sub-0.5B models with a normalized score of 62.3. During testing, it surpassed competing small-scale AI models from Liquid AI and Hugging Face, outperforming all tested 0.5B-class models across 16 out of 17 benchmarks.

The model also leads across document understanding, OCR, visual perception, reasoning and reliability tasks, Tether said.

It achieved a 4.6% relative advantage in visual perception and a 7.4% advantage in reasoning compared with other models in the same size class. It also outperformed significantly larger models in instruction-following benchmarks, including FastVLM-0.5B, Qwen3.5-0.8B and InternVL3.5-1B.

VisionPsy-Nano is available under the Apache 2.0 license, allowing developers and researchers to use them through multiple deployment options, including Hugging Face Transformers, llama.cpp-based mobile inference and vLLM-powered server deployments. Tether also released evaluation configurations to enable benchmark reproduction.

Through QVAC, Tether Data is pursuing decentralized AI systems designed around privacy, user control and device-based intelligence rather than large-scale data centre dependence.