VisionPsy-Nano-460M

VisionPsy-Nano-460M is QVAC's compact open-weight vision-language model for single-image understanding on phones and edge devices, with an 8,192-token context window.

Platform: Replicate

Vision Language ModelMultimodalOn-Device AIEdge AIOpen WeightQVAC

Local CPU or CUDA through Transformers / on-device GGUF runtime; exact device support depends on backend

Open-weight Apache 2.0 model; local inference cost depends on device, runtime, quantization, and electricity

Commercial

🚀Function Overview

VisionPsy-Nano-460M is a roughly 460M-parameter image-to-text model built on nanoVLM with a SigLIP2 vision encoder, SmolLM2 language backbone, 8K context, and Apache 2.0 weights.

Key Features

Official Hugging Face model ID qvac/VisionPsy-Nano-460M
Roughly 460M parameters for edge and on-device deployment
8,192-token context window
Native 512x512 image processing with tiling for larger images
Single-image question answering and document, chart, diagram, and scene-text understanding
Apache License 2.0 open weights
Direct Transformers loading with trust_remote_code
Separate VisionPsy-Nano-460M-Flash checkpoint for lower first-token latency
Official GGUF variants for on-device runtimes
Primarily English and not intended for safety-critical automated decisions

Use Cases

•Private or offline visual question answering
•Reading charts, diagrams, documents, and scene text on an edge device
•Camera and image assistants constrained by phone memory or battery
•Domain fine-tuning for a focused visual workflow
•Comparing full-quality and Flash checkpoints on one device
•Testing compact VLM quality, latency, and memory tradeoffs

⚙️Input Parameters

image

One image per query. The official card recommends single-image inputs and native 512x512 processing with tiling for larger images.

text

string

A plain-text question or instruction about the image.

max_new_tokens

integer

Maximum generated response length for the local Transformers generation call.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "qvac/VisionPsy-Nano-460M",
  "image": "chart.png",
  "text": "What changed most between the two periods, and which labels support that conclusion?",
  "max_new_tokens": 160
}

Output Results

The model returns a text answer grounded in the supplied chart. Verify numbers and labels before using the result in a high-stakes workflow.

Quick Actions

Use NowView Documentation

Technical Specifications

Hardware Type: Local CPU or CUDA through Transformers / on-device GGUF runtime; exact device support depends on backend
Commercial Use: Supported
Pricing: Open-weight Apache 2.0 model; local inference cost depends on device, runtime, quantization, and electricity
Platform: Replicate

Related Keywords

VisionPsy-NanoVisionPsy-Nano-460Mqvac VisionPsyon-device vision language modeledge VLMsmall vision language modelVisionPsy Nano FlashVisionPsy GGUFimage question answering model

Related Models

Bielik 1.5B v3 Instruct

Bielik-1.5B-v3-Instruct is a generative text model featuring 1.6 billion parameters. It is result of collaboration between the open-science/open-souce project SpeakLeash and the High Performance Computing (HPC)

Cordia-A6 Text Generation Model

A model for generating text sequences based on input prompts and adjustable parameters.

Claude Sonnet 4

Claude Sonnet 4 is a significant upgrade to 3.7, delivering superior coding and reasoning while responding more precisely to your instructions