A new compact vision-language model from Tether Data is challenging the assumption that powerful AI requires cloud computing. VisionPsy-Nano, a 460-million-parameter model designed specifically for smartphones and edge devices, has achieved the highest performance score among all comparable compact models, outperforming competitors from Liquid AI and Hugging Face while running entirely on-device.

What Makes This Model Different From Cloud-Based AI?

Vision-language models (VLMs) are AI systems that understand both images and text, answering questions about photos, reading documents, and analyzing visual scenes. Traditionally, this kind of sophisticated analysis required sending images to distant data centers for processing. VisionPsy-Nano flips that script by delivering the same capabilities directly on your phone, keeping images private and eliminating the delay of cloud uploads.

The model achieved an overall normalized score of 62.3 on industry benchmarks, compared to 59.6 for Liquid AI's LFM2.5-VL-450M, 52.5 for Hugging Face's SmolVLM2-500M, and 54.9 for Tether's own base model. This means VisionPsy-Nano outperformed all competing models in its size category on 16 out of 17 tested benchmarks.

How Does It Handle Real-World Tasks on Your Phone?

Tether Data released two versions of VisionPsy-Nano to meet different needs. The standard version prioritizes accuracy across all vision tasks, while a faster variant called VisionPsy-Nano-Flash trades a tiny amount of quality (retaining 99% of the full model's performance) for dramatically faster response times. On a Pixel 9, Galaxy S23, Galaxy S25 Ultra, and iPhone 15, the Flash version generates its first response 19 to 36 times faster than competing models, depending on the device.

The model excels at four distinct types of visual understanding tasks:

• Document Understanding: Extracts text and structural insights from complex backgrounds, flowcharts, financial reports, and infographics with high accuracy.
• Visual Perception: Analyzes scenes and spatial layouts with a 4.6% performance advantage over the next-best model in its size class.
• Reasoning and Knowledge: Performs visual reasoning tasks with a 7.4% relative margin over competing compact models.
• Instruction Following: Follows user commands reliably, beating models 1.6 to 2.3 times its size on instruction-following benchmarks.

Why Should Developers Care About On-Device Inference?

Tether Data released VisionPsy-Nano as open-source software under the Apache 2.0 license, meaning developers can freely use and modify it. The company provided three different ways to run the model: through standard machine learning libraries, as a quantized version optimized for mobile phones via llama.cpp, and as a production-ready backend for servers using vLLM.

This approach addresses a growing tension in AI development. Cloud-based AI offers powerful computing but raises privacy concerns, requires constant internet connectivity, and incurs ongoing costs. On-device models keep data local, work offline, and eliminate latency, but historically have been too small or slow to handle complex tasks like image understanding. VisionPsy-Nano demonstrates that this trade-off is becoming less necessary.

"The technology is ready to counter the industry's obsession with data center centralization right now. 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," said Paolo Ardoino, CEO of Tether.

Paolo Ardoino, CEO of Tether

How to Deploy VisionPsy-Nano on Your Device

• Full-Precision Quality: Access the model through Hugging Face Transformers, the standard library for running machine learning models, if you need maximum accuracy and have sufficient device memory.
• Mobile Deployment: Use quantized GGUF checkpoints designed to run on phones through llama.cpp with a single command, ideal for smartphones and tablets with limited processing power.
• Server Inference: Deploy production-ready backend inference powered by vLLM for high-throughput applications that need to process many images simultaneously.

All evaluation configurations and benchmark setups are included with the release, allowing researchers to reproduce the results and verify the performance claims independently.

The release reflects a broader shift in AI development toward efficiency and decentralization. Rather than concentrating computing power in massive data centers, researchers are increasingly building smaller, smarter models that run on the devices people already own. For users, this means faster responses, better privacy, and AI that works even without an internet connection. For developers, it opens new possibilities for applications that were previously impractical on mobile devices.