Tether Data has released an open-source vision-language model designed to run directly on smartphones, in a move its creators say shows advanced image-focused AI no longer has to depend on cloud computing.
The system, called VisionPsy-Nano, has 460 million parameters and has been built to analyse images, documents and charts entirely on a user’s device. Developed by Tether Data’s AI research division, QVAC, the model was made available on Wednesday and can operate without an internet connection, with both inputs and responses processed locally.
According to QVAC, VisionPsy-Nano has achieved the highest overall performance of any vision-language model with fewer than 500 million parameters. In tests across 17 different benchmarks, it outscored rival models in 16 of them.
On a normalized scoring scale, VisionPsy-Nano recorded 62.3 points. That compares with 59.6 for Liquid AI’s LFM2.5-VL-450M, 52.5 for Hugging Face’s SmolVLM2-500M and 54.9 for QVAC’s own earlier nanoVLM-460M-8k base model, underlining what the company says is a significant step forward in capability at this compact size.
Two builds aimed at accuracy and speed
Tether Data is offering the model in two configurations. The standard version is tuned for maximum accuracy, while a second variant, named Flash, trades a small degree of quality for much faster response times.
QVAC says Flash maintains about 99% of the full model’s performance but delivers its first output considerably more quickly than existing competitors. On Android devices, Flash is claimed to generate an initial response up to 23 times faster than SmolVLM2-500M, and up to 36 times faster on an iPhone 15.
The company argues that this initial latency is critical on mobile hardware. Even strong benchmark scores can feel academic, it says, if users are left waiting too long for the device to process an image. Flash has been engineered specifically to cut that delay and make on-device vision AI feel more responsive in everyday use.
Document understanding and visual reasoning
Beyond general image recognition, VisionPsy-Nano is also geared towards document analysis and optical character recognition. QVAC says the model can extract text and structural information from financial reports, flowcharts and infographics, aiming to keep accuracy high even on dense or complex layouts that typically challenge smaller systems.
On visual reasoning benchmarks, QVAC reports that the model outperformed similarly sized rivals by an average margin of 7.4%. It also says VisionPsy-Nano surpassed models more than twice its size on the MM-IFEval and POPE evaluations, including releases from Qwen and InternVL.
Running such tasks directly on phones and other consumer hardware imposes tight constraints around memory, heat, battery life and processing power. Smaller models can often manage plain text, but many lose precision when confronted with intricate tables, charts or scanned paperwork. QVAC’s internal results suggest VisionPsy-Nano has retained much of that capability while remaining compact enough for local deployment.
The company also highlights the privacy and resilience benefits of processing data on-device. Because documents do not need to be uploaded to remote servers, sensitive material can remain on the user’s hardware, and the model can continue working even without a network connection.
Access, licensing and research
Developers can use VisionPsy-Nano through several popular tools and frameworks. QVAC has made the model available via Hugging Face Transformers, as a quantized GGUF build for llama.cpp, and with a vLLM backend for scenarios requiring higher-volume server deployment.
To support independent verification, QVAC has also published its benchmark configurations through VLMEvalKit, enabling researchers and third parties to reproduce its tests. Both the standard and Flash versions are released under the Apache 2.0 licence, which permits commercial use, modification and redistribution.
Tether CEO Paolo Ardoino said the launch fits into the company’s broader strategy of moving AI workloads closer to the user.
“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,” Ardoino said.
Part of a wider AI expansion
VisionPsy-Nano is the latest in a sequence of QVAC releases dating back to October 2025. Previous work has included synthetic training datasets, a cross-platform software development kit and medical AI models tailored for smartphones and wearable devices.
For software developers, the new model offers a way to deliver offline image and document analysis without relying on usage-based API services. Businesses can keep confidential files on local devices while applying AI tools, and consumers can access features even when they have no signal. Academic and industry researchers are also able to assess QVAC’s claims independently using the published evaluation settings.
Tether has been funding its push into artificial intelligence with profits from its USDT stablecoin operation. Alongside its AI work, the company has been investing in broader technology areas including AI infrastructure, biotechnology and robotics. Its backing includes a Series C investment in NEURA Robotics, valued at up to $1.4bn.
The company’s latest move underlines a growing trend in the AI sector towards smaller, more efficient models that can run directly on personal devices, reducing dependence on large-scale cloud data centres while promising faster responses and greater privacy for users.
