Multimodal Overview

For supported models, send messages[].content as an OpenAI-compatible content-part array instead of a plain string. Currently accepted parts:

text  image_url  input_audio 

input_file

When should you use it?

Note

Use this structure only with models whose detail page shows the related capabilities. A text-only model will reject image_url with a 400.

Read a single model's modalities from model discovery; the endpoint needs no authentication:

curl -s https://llmtr.com/v1/models/google/gemini-3.6-flash | jq '.architecture'

To list every model that supports a specific operation (image generation, embeddings, TTS, …) filter the catalog on supported_operations:

# All image-generating models
curl -s https://llmtr.com/v1/models \
  | jq -r '.data[] | select(.supported_operations[] == "IMAGES_GENERATIONS").id'

# Embedding models
curl -s https://llmtr.com/v1/models \
  | jq -r '.data[] | select(.supported_operations[] == "EMBEDDINGS").id'

# Text-to-speech models
curl -s https://llmtr.com/v1/models \
  | jq -r '.data[] | select(.supported_operations[] == "AUDIO_SPEECH").id'

When a model is sent to the wrong endpoint, the 400 unsupported_operation response includes error.details.supported_endpoints and error.details.suggested_endpoint to point you at the correct route. See Errors.

Content-part shape

{
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "What's in this image?" },
        {
          "type": "image_url",
          "image_url": { "url": "https://..." }
        }
      ]
    }
  ]
}

Limits and notes

• Media is sent through the JSON body. Remote URLs are safer than inline base64.
• Keep inline base64 audio clips short (< 1 MB recommended).
• For large files, PDFs, video, or reusable media use the Files API.

Sub-pages

• Image input
• Audio input
• Files API