Tether Data Pushes AI Off the Cloud With New 460M Parameter Vision Model

Tether Data has open-sourced a 460-million-parameter vision-language model built to run directly on smartphones instead of relying on cloud servers.

QVAC says Tether AI Research’s model recorded the highest overall score among vision-language systems with fewer than 500 million parameters. Across 17 benchmarks, it beat competing models on 16.

The model posted a normalized score of 62.3, compared with 59.6 for Liquid AI’s LFM2.5-VL-450M and 52.5 for Hugging Face’s SmolVLM2-500M. QVAC’s earlier nanoVLM-460M-8k base model scored 54.9.

That response time matters on mobile hardware. A compact model can score well in testing but still feel impractical if users must wait while the device processes an image. Flash is designed to reduce that initial delay.

QVAC says the model beat similarly sized competitors by an average of 7.4% on visual reasoning tests. It also outperformed models more than twice its size on MM-IFEval and POPE, including releases from Qwen and InternVL.

Moving image processing from a data center to a phone introduces strict limits around memory, heat, battery use, and computing power. Compact models often handle plain text but lose accuracy when reading dense charts, tables, or scanned documents.

QVAC’s benchmark results suggest the model preserved much of that capability while remaining small enough for consumer devices. Local processing also means documents do not need to be uploaded, and the software can continue working without a network connection.

Developers can access the model through Hugging Face Transformers, a quantized GGUF version for llama.cpp or a vLLM backend for higher-volume server use.

QVAC also published its benchmark configurations through VLMEvalKit, allowing outside researchers to repeat the tests. Both versions use the Apache 2.0 license, permitting commercial use, modification, and redistribution.

“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.

The model follows several QVAC releases dating to October 2025, including synthetic training datasets, a cross-platform software development kit, and medical models designed for phones and wearable devices.

For developers, the release offers offline image analysis without usage-based API costs. Businesses can keep sensitive documents on the device, while consumers can use AI features without a signal. Researchers can independently test the company’s performance claims using the published configurations.