# Tether releases 460M-parameter VisionPsy-Nano models for on-device vision AI > Source: > Published: 2026-07-29 18:21:48+00:00 Tether Data's QVAC research group released two 460M-parameter vision-language models on July 29th, giving developers open weights and mobile inference code for applications that need to interpret images without sending them to a cloud service. QVAC [announced VisionPsy-Nano in a thread on X](https://x.com/qvac/status/2082439175891284476?s=46) and published the [model collection](https://huggingface.co/collections/qvac/visionpsy) under the Apache 2.0 license. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) [https://x.com/qvac/status/2082439175891284476?s=46](https://x.com/qvac/status/2082439175891284476?s=46) The release extends [Tether CEO Paolo Ardoino's](https://tether.io/news/tether-appoints-paolo-ardoino-as-ceo/) push to turn the stablecoin issuer into a builder of peer-to-peer computing infrastructure. Ardoino joined Bitfinex in 2014, became Tether's [chief technology officer](/article/rob-witoff-joins-coinbase-as-cto) in 2017 and took over as CEO in December 2023. When Tether introduced QVAC in May 2025, he summarized the strategy plainly: "If you need an API key to use your AI, it isn't truly yours." ([tether.io](https://tether.io/news/tether-appoints-paolo-ardoino-as-ceo/)) The [technical post](https://huggingface.co/blog/qvac/visionpsy), credited to a 10-person research group including [Khurram Azeem Hashmi](https://huggingface.co/Khashmi) and [Mathias Buus](https://huggingface.co/mafintosh), presents two variants. [VisionPsy-Nano-460M](https://huggingface.co/qvac/VisionPsy-Nano-460M) prioritizes quality, while [VisionPsy-Nano-460M-Flash](https://huggingface.co/qvac/VisionPsy-Nano-460M-Flash) reduces the visual tokens processed for each image to cut latency and memory use. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) ### A smaller visual pipeline VisionPsy-Nano builds on the nanoVLM architecture, pairing a SigLIP2 image encoder with a 360M-parameter SmolLM2 language backbone and a lightweight connector. Both variants support an 8,192-token context and are designed around one image per request. The quality model can produce 1,088 visual tokens for a 512x512 image; Flash processes the same resolution with 64 visual tokens by preserving the image's native scale instead of enlarging it before tiling. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) QVAC measured Flash on a Pixel 9, Galaxy S23, Galaxy S25 Ultra and iPhone 15 using four-bit GGUF builds. In those tests, Flash reached its first output token 11.9x to 25.1x faster than nanoVLM-460M and SmolVLM2-500M on a common runtime. It also reached the first token 1.3x to 2.4x faster than Liquid AI's LFM2.5-VL-450M and 2.3x to 3.5x faster than Qwen3.5-0.8B when each phone used its fastest tested backend. Those are QVAC's measurements rather than independently reproduced results. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) Developers can load the full-precision checkpoints through Transformers, use vLLM for server inference or deploy quantized GGUF versions through llama.cpp. QVAC also released a [GitHub repository with inference scripts](https://github.com/tether-ai-research/qvac-visionpsy-nano) for all three paths. The broader [QVAC SDK](https://github.com/tetherto/qvac) uses JavaScript and TypeScript interfaces for local AI across Linux, macOS, Windows, Android and iOS. ([github.com](https://github.com/tetherto/qvac)) ### The benchmark lead comes with caveats QVAC reports a normalized aggregate score of 62.3 for VisionPsy-Nano-460M, ahead of LFM2.5-VL-450M at 59.6, nanoVLM-460M-8k at 54.9 and SmolVLM2-500M at 52.5. QVAC says VisionPsy-Nano won 15 of 17 benchmark rows outright against that roughly 0.5B-parameter group and compared favorably on a 16th. It led the grouped scores for document understanding and OCR, visual perception, reasoning and knowledge, and instruction following and reliability. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) The comparisons were run in-house with a single VLMEvalKit harness. QVAC used Qwen3.6-27B to judge open-ended answers, a method intended to avoid penalizing responses for formatting differences such as "12%" versus "12 percent." At launch, the changes for VisionPsy model support and the additional judge-based scoring remained open pull requests in VLMEvalKit and had not received maintainer reviews. QVAC published its configurations so researchers can reproduce the setup by applying those patches. ([github.com](https://github.com/open-compass/VLMEvalKit/pull/1613)) The parameter-class comparison also has a defined boundary. Larger models still won the raw 17-benchmark average: Qwen3.5-0.8B scored 66.1 and InternVL3.5-1B scored 65.9, compared with VisionPsy-Nano's 62.3. VisionPsy-Nano beat those larger models on three complete benchmarks covering science questions, visual instruction following and hallucination resistance. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) ### Tether's local AI wedge QVAC targets camera questions, receipts, forms, charts, diagrams, scene text and lightweight visual instructions. The small parameter count creates hard limits: QVAC recommends domain-specific fine-tuning, supports English primarily and warns that the models can hallucinate, miscount or struggle with dense documents and multi-step mathematics. QVAC says VisionPsy-Nano should not be used for safety-critical automated decisions. ([huggingface.co](https://huggingface.co/blog/qvac/visionpsy)) Tether first announced QVAC in May 2025 and publicly launched it at the Plan B Forum in Lugano that October. The developer SDK followed in April 2026. VisionPsy-Nano gives that local-first thesis a model that developers can inspect and deploy, rather than another layer that depends on a third-party model API. Its adoption will depend on whether independent evaluations confirm QVAC's benchmark lead and whether the latency gains hold inside complete mobile applications, where image capture, preprocessing and device-specific integration add costs beyond model inference. ([qvac.tether.io](https://qvac.tether.io/blog/the-future-of-ai-is-local/))