Tether Data Releases Open Source Edge Visual Language Model VisionPsy-Nano
Tether Data's AI research project QVAC has open-sourced VisionPsy-Nano, a compact visual language model designed for edge and on-device deployment, with approximately 460 million parameters. This model brings multi-modal understanding capabilities, previously reliant on cloud infrastructure, directly to mobile devices.
Among all evaluated edge visual language models with fewer than 500 million parameters, this model achieved the highest overall normalized score of 62.3, outperforming competitors like Liquid AI and Hugging Face in 16 out of 17 benchmark tests. A Flash version optimized for low latency retains about 99% of the full model quality while significantly improving inference speed; both versions are open-sourced under the Apache 2.0 license.
Event-driven local-first AI resources are concentrating on edge devices, benefiting mobile and privacy-sensitive applications while putting pressure on service providers reliant on cloud inference.
Source: Public Information
ABAB AI Insight
Tether continues to build a decentralized AI through the QVAC project, extending its historical path from stablecoin infrastructure to locally runnable models. VisionPsy-Nano maintains its positioning of "infinite intelligence, local hardware, peer-to-peer," achieving leading capabilities in document OCR, visual perception, reasoning, and instruction following with fewer than 500 million parameters.
On the capital path, open-sourcing Apache 2.0 weights and providing multiple deployment options such as Transformers, llama.cpp, and vLLM aims to bypass centralized data centers, proving the feasibility of efficient local AI, allowing powerful private models to run directly on users' existing devices, while reserving expansion for future DeFi and cross-chain scenarios.
Similar cases can be seen in lightweight multi-modal attempts like Apple FastVLM and Hugging Face SmolVLM. Currently, edge visual language models are transitioning from cloud dependency to device-native operation, with the low-latency Flash variant becoming a key differentiator.
This essentially represents a technological substitution: compact open-source models replacing cloud-based multi-modal APIs, with mechanisms based on parameter compression and visual token optimization, transforming privacy and latency advantages into the default choice for edge deployment.
ABAB News · Cognitive Laws
Smaller parameters do not equal weaker capabilities
Local-first determines privacy boundaries
Open-source weights accelerate edge AI implementation