The QVAC research initiative by Tether Data designed this model to handle complex visual tasks—ranging from OCR and document analysis to spatial reasoning—without needing external servers. By releasing both a standard version and a 'Flash' variant tuned for latency, the company is targeting developers who require real-world responsiveness. The Flash edition reportedly achieves first-token generation up to 36 times faster than previous benchmarks on an iPhone 15, while maintaining 99% of the full model's accuracy.
In section Releases
Tether Releases VisionPsy-Nano to Bring High-End AI to Smartphones
With a normalized score of 62.3, Tether Data’s new 460-million-parameter VisionPsy-Nano model has claimed the top spot for performance among sub-0.5B vision-language models. The open-source release aims to shift multimodal AI away from cloud-dependent infrastructure and directly onto everyday mobile hardware like the iPhone 15 and Galaxy S25 Ultra.
VisionPsy-Nano outperformed competitors including Liquid AI’s LFM2.5-VL-450M and Hugging Face’s SmolVLM2-500M across 16 of 17 industry benchmarks. Paolo Ardoino, CEO of Tether, stated that the release is intended to counter the trend of data center centralization. By providing open-weight access under an Apache 2.0 license, the project allows for deployment via llama.cpp or vLLM, effectively placing high-performance multimodal capabilities into the hands of local device users.
Comments (0)
No comments yet. Be the first!