VisionPsy-Nano
VisionPsy-Nano is a QVAC vision-language model family released on July 29, 2026. The family includes the quality-focused `qvac/VisionPsy-Nano-460M` and the latency-focused `qvac/VisionPsy-Nano-460M-Flash`; both are Apache-2.0 open-weight models built for image question answering, document and chart reading, and other on-device visual tasks.
VisionPsy-Nano makes the quality-versus-latency choice explicit inside one small-model family. The full model prioritizes benchmark quality, while Flash reduces visual-token count for faster first-token latency and lower memory use. That makes the family useful when a phone, embedded device, privacy requirement, or offline workflow rules out a cloud-scale VLM.
QVAC's official launch post and Hugging Face collection establish the release, model IDs, open weights, and Apache 2.0 license. The official model cards document the 8,192-token context, single-image design, architecture, deployment requirements, benchmark harness, and Flash tradeoffs. Performance claims remain vendor-reported because the launch is new and independent reproduction is not yet mature.
- Single-image question answering on a phone, laptop, or edge device.
- Document, chart, diagram, and scene-text understanding without a cloud model.
- Private or offline visual assistants with constrained memory and power.
- Comparing full-quality and low-latency checkpoints on the same image workload.
`VisionPsy-Nano-460M` is the quality-focused checkpoint. `VisionPsy-Nano-460M-Flash` keeps the same family architecture but reduces visual tokens, including 64 tokens for a native 512×512 image, to lower time to first token and memory use. Use the full model when OCR and fine visual detail matter most; test Flash when responsiveness, battery, or RAM is the binding constraint.
The official card targets visual question answering, document, chart, diagram, and scene-text understanding. It supports an 8,192-token context and native 512×512 image processing with tiling for larger images. The full model is designed for one image per query, primarily in English, and the card warns that compact size can still produce hallucinations, counting errors, and weaker results on dense documents or long multi-step mathematics.
QVAC reports one in-house VLMEvalKit harness across 17 public benchmarks and uses Qwen3.6-27B as an LLM judge for some open-ended tasks. The model card also states that several evaluation changes are still under review upstream. Those details make the results inspectable, but they do not make them independent; validate the exact device, backend, images, latency metric, and answer-quality threshold that matter to your application.
Source confidence
QVAC
QVAC on Hugging Face
QVAC on Hugging Face
QVAC on Hugging Face
QVAC on Hugging Face
VisionPsy-Nano FAQ
Common questions about VisionPsy-Nano.
Is VisionPsy-Nano open source or open weight?+
The model weights are public under the Apache License 2.0. QVAC publishes the full and Flash checkpoints plus GGUF variants on Hugging Face. Review the model card and any dataset licenses separately before training or redistributing a derivative.
What is the difference between VisionPsy-Nano and VisionPsy-Nano Flash?+
The full checkpoint prioritizes visual quality. Flash reduces visual tokens to improve time to first token and memory use, with a documented quality tradeoff on some OCR and fine-detail tasks. Benchmark both on the same images and device before choosing.
Can VisionPsy-Nano run on a phone?+
QVAC publishes phone measurements and GGUF variants for on-device use, but speed depends on the phone, backend, image resolution, quantization, and whether CPU or GPU acceleration is available. Treat the launch numbers as vendor-reported until reproduced on your hardware.
Does VisionPsy-Nano support multiple images?+
The full model card describes a single-image design and recommends one image per query. Multi-image prompts are outside its intended use, so use a different VLM or redesign the workflow if cross-image reasoning is required.