AI Concepts

What are multimodal models?

Multimodal models accept and produce more than one data type — typically text plus images, sometimes audio or video — within a single request. A vision-capable chat model can read screenshots, diagrams, charts or scanned documents and reason over them alongside text. Plugsky exposes multimodal chat through the same OpenAI-compatible API; image generation is not yet a live endpoint.

Key facts

DefinitionModels that process multiple input or output modalities in one request
Why it mattersDocuments, screenshots and diagrams no longer need separate OCR pipelines
Plugsky supportVision-capable chat models are reachable through the OpenAI-compatible chat completions API
Typical inputsText plus images; audio and video support varies by model
Use casesDocument Q&A, UI testing, chart reading, accessibility and content moderation review
StatusMultimodal chat is live; image generation is a coming-soon endpoint
Models30+ models behind one API, with capability documented per model
Cost driverImages consume tokens based on resolution and count, so size inputs deliberately

TL;DR

  • Multimodal means more than text in one request — usually text plus images.
  • Vision models remove separate OCR steps for many document workflows.
  • Image token cost scales with resolution, so downscale where possible.
  • Capability differs per model — check the catalogue before committing.
  • Plugsky serves multimodal chat live; image generation is coming soon.

How it works, step by step

  1. Identify where images carry information today: screenshots, scans, charts or photos.
  2. Check which models in the catalogue accept image input for your task type.
  3. Encode images per the API format and send them alongside the text prompt.
  4. Downscale images to the minimum resolution that still answers the question.
  5. Validate outputs against extracted ground truth on a labeled sample.
  6. Fall back to OCR plus a text model where vision quality or cost does not fit.
1Identify whereimages carryinformation today:2Check which modelsin the catalogueaccept image input3Encode images perthe API format andsend them alongside4Downscale images tothe minimumresolution that5Validate outputsagainst extractedground truth on a6Fall back to OCRplus a text modelwhere vision

Try it yourself

Open the best AI model selector →

What multimodal means in practice

A text-only model sees characters. A multimodal model receives a prompt built from several modalities — commonly text and images — and reasons across them. In practice this means you can send a screenshot of a broken UI and ask what is wrong, or a scanned invoice and ask for structured fields. The model handles the perceptual step that previously required a separate OCR or computer-vision pipeline.

When to use multimodal models

  • Document understanding: scanned forms, invoices and PDFs where layout matters.
  • UI and QA: screenshot-based regression triage and accessibility review.
  • Data extraction: reading charts, tables and gauges from images.
  • Support: customers attach screenshots and expect an answer in one turn.
  • Content review: flagging policy issues in uploaded images.

If the input is already clean text, a text model is cheaper and faster — use vision only when the image itself carries the information.

Common mistakes

  • Sending full-resolution photos when a downscaled image answers the question, inflating token cost.
  • Assuming every model in the catalogue accepts images — capability varies and must be checked per model.
  • Trusting extracted numbers without a verification step on financial or medical data.
  • Mixing many images into one request and losing track of which one supports the answer.
  • Ignoring privacy rules when images contain faces, identity documents or customer data.

How Plugsky implements it

Plugsky serves multimodal chat through the OpenAI-compatible chat completions endpoint, so images are passed in the same message format used by the OpenAI SDK — change the base URL and your existing code path works. Vision-capable models are listed in the catalogue with their capabilities. Embeddings and RAG remain live for text pipelines that complement vision, while image generation is documented as a coming-soon endpoint. For regulated images, private deployment options keep traffic inside your network.

Honest comparison

ApproachMultimodal modelOCR plus text modelCustom computer vision
Setup effortLow — one API callMedium — OCR pipelineHigh — training and serving
Layout understandingStrong on documents and UIsVaries by OCR engineTask-specific
FlexibilityGeneral-purpose promptsFixed extraction schemaNarrow
CostImage tokens per requestOCR plus text tokensGPU hosting
Best forMixed visual and text reasoningHigh-volume structured formsNarrow repetitive checks

Frequently asked questions

What are multimodal models?

AI models that process more than one data type in a single request — most commonly text and images, sometimes audio or video — and reason across the combined input.

Which plugsky models support images?

Vision support varies by model, so check the model catalogue for current capabilities before building a workflow that depends on image input.

Do multimodal models replace OCR?

For many document and screenshot workflows, yes — the model reads the image directly. For very high-volume, fixed-schema extraction, a dedicated OCR pipeline can still be cheaper.

How does image cost work?

Images are converted to tokens based on resolution and count, so they consume part of the context window. Downscale images to the smallest size that still contains the information.

Can Plugsky generate images?

Image generation is documented as a coming-soon endpoint. Multimodal understanding — text plus images in chat — is live today.

Is multimodal chat OpenAI-compatible?

Yes. Images are sent in the same message format as the OpenAI API, so existing SDK code works after changing the base URL.