Which Plugsky model is best for vision tasks?

Key facts

Fast vision	plugsky-vision-fast for screenshots, captioning and quick triage
Vision plus tools	plugsky-qwen-vl for vision-language tasks that use function calling
Compact multimodal	plugsky-gemma-4 and the Gemma 3 nano tiers for extraction
General chat	plugsky-plus adds vision to everyday chat flows
API shape	Images are sent inside the chat messages array
Not included	Image generation and editing endpoints — check the docs for current status
Free plan	plugsky-micro and plugsky-lite cover text work; no card required
Product status	Live

TL;DR

• plugsky-vision-fast is the default for screenshots and high-volume visual triage.
• plugsky-qwen-vl adds tool calling when vision must trigger actions.
• plugsky-gemma-4 and Gemma 3 nano tiers suit compact multimodal extraction.
• General tiers such as plugsky-plus bring images into normal chat flows.
• Image generation is not part of the vision chat models — verify endpoint status.

How it works, step by step

1. List your image tasks: triage, captioning, form extraction, chart reading or visual QA.
2. Test plugsky-vision-fast on a sample set of real images first.
3. Add plugsky-qwen-vl when the workflow must call tools after seeing an image.
4. Use a compact multimodal tier for high-volume, narrow extraction jobs.
5. Escalate images that need deep reasoning to a stronger vision-capable tier.
6. Validate image handling end to end: encoding, size limits, multiple images per request.
7. Score accuracy per task and route each image type to the cheapest passing model.

Try it yourself

Open the AI model selector →

What vision on the API means

Vision models accept images inside the chat messages array, alongside text. You send an image reference or encoded payload, add a question, and receive a normal completion. Typical uses are screenshot triage, receipt and form extraction, chart reading, product photo captioning and visual quality checks.

This is input-side multimodality. The vision chat models do not generate or edit images; those are separate endpoints, and you should check the docs for their current status before committing a product roadmap to them. Capability flags per model are listed on the live model card at /models.

Choosing a vision tier

plugsky-gemma-4 is the compact multimodal option for high-volume extraction. plugsky-vision-fast targets fast image and screenshot triage, while plugsky-qwen-vl is the vision-language tier for workflows that combine images with tool calling. When an image needs genuine reasoning — comparing diagrams, interpreting ambiguous charts — escalate to a stronger vision-capable tier rather than retrying the fast one.

Vision models pair naturally with retrieval: extract text from images once, embed it with plugsky-embed, and answer future questions from the index instead of re-processing images every time.

Limits and evaluation

Image inputs consume context, so resolution and image count affect both latency and cost. Resize before sending where possible, send only the frames or pages the model needs, and prefer one well-chosen image over five redundant ones.

• Build a labelled image set per task and score accuracy, not just fluency.
• Include hard cases: rotated scans, handwriting, low light and screenshots with UI chrome.
• Test multiple-image requests if your workflow compares documents.
• Log which image produced each answer so failures can be reproduced.

Route by image type, keep a stronger fallback for low-confidence results, and re-run the evaluation when the model card changes.

Honest comparison

Vision task	Fast vision tier	Vision plus tools	Compact multimodal
Screenshots and UI triage	Best fit	Strong	Good
Documents and forms	Good	Strong	Best fit for compact extraction
Charts and diagrams	Limited	Strong	Limited
Captioning at volume	Best fit	Strong	Strong
Tool use after seeing an image	Not the focus	Best fit	Limited

Frequently asked questions

Which Plugsky model should I use for image tasks?

Start with plugsky-vision-fast for triage and captioning, plugsky-qwen-vl when the workflow also calls tools, and plugsky-gemma-4 for compact high-volume extraction. Evaluate on your own images before standardising.

Can Plugsky models generate images?

The vision chat models accept images as input; image generation and editing are separate endpoints. Check the docs for their current status before building them into a product.

How are images sent to the API?

Images are included in the chat messages array alongside text, using the standard OpenAI-compatible chat completions shape. The docs cover accepted formats and size guidance.

Do vision calls cost more?

Image inputs consume context, so they affect usage more than a short text prompt. Self-serve plans are flat monthly with fair-use usage — see the live pricing page for plan details.

Can I use vision on the free plan?

The free plan covers plugsky-micro and plugsky-lite for text with no card required. Vision tiers are paid; use the 14-day full-access trial to evaluate them.

How do I evaluate a vision model?

Build a labelled set per task, include rotated and low-quality scans, and measure exact-match or rubric-scored accuracy. Review failures to see whether resolution, prompt or model choice caused them.

Should I combine vision with RAG?

Often yes. Extract text from images once, embed it with plugsky-embed, and answer future questions from the index — cheaper and more consistent than re-sending images each time.