Image Input (Vision)

LLMAPI supports image input for vision-capable language models (VLMs) through the OpenAI-compatible /v1/chat/completions endpoint. Use the same message structure as the OpenAI vision guide: a user message can contain an array of text parts and image_url parts.

Default to low detail

Set detail: "low" unless the model must inspect small text, dense diagrams, or fine visual details. Low detail is faster and reduces image-token usage, which keeps managed-model cost lower.

Supported models

The following current LLMAPI models have image in their configured input modalities:

Region	Model	Provider	Cost guidance	Notes
RoW	qwen3.6-35b-a3b	Self-hosted CAIP	Free	Recommended when self-hosted data sovereignty and zero usage cost are important.
RoW	qwen3.6-35b-a3b-think	Self-hosted CAIP	Free	Same vision capability with extended thinking enabled. Both self-hosted vision models also accept video input.
RoW	gpt-4o	Azure OpenAI	Managed model pricing	Strong managed VLM; use only when higher quality is required.
RoW	gpt-4o-mini	Azure OpenAI	Managed model pricing	Lower-cost managed option for simple visual understanding.
RoW	gpt-41, gpt-41-mini, gpt-41-nano	Azure OpenAI	Managed model pricing	Use the smallest model that meets quality requirements.
RoW	gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-2, gpt-5.5, gpt-5-pro, gpt-5-codex	Azure OpenAI	Managed model pricing	Multimodal GPT models; reserve larger variants for difficult tasks.
RoW	o3, o3-pro	Azure OpenAI	Managed model pricing	Reasoning models for complex visual reasoning tasks.
RoW	claude-3-haiku, claude-37-sonnet, claude-4-sonnet, claude-haiku-4.5, claude-sonnet-4.6, claude-opus-4.5, claude-opus-4.6	AWS Bedrock (Anthropic)	Contact CAIP	Claude models support image input; legacy Claude model names forward to newer Claude deployments.
RoW	nova-lite, nova-pro	AWS Bedrock (Amazon Nova)	Contact CAIP	Nova models support text, image, and video input.
China	qwen3-vl-plus, qwen3-vl-flash, qwen-vl-max	Alibaba Cloud	Contact CAIP	Qwen VL models for image and video understanding.
China	MiniMax-M2.5	Alibaba Cloud	Managed model pricing	Multimodal chat model.

Models exposed through another endpoint, for example Responses API models such as gpt-5-pro, gpt-5-codex, and o3-pro, support image input but use that endpoint's request format instead of the chat-completions examples below.

qwen3-chat does not support image input. If you send images to a model that does not support vision, LLMAPI may fall back to another model. Always check the model field in the response, or set disable_fallbacks: true if you want an error instead of fallback routing.

Request format

Use the standard chat completions endpoint:

POST https://llm.api.caip.bmw.cloud/v1/chat/completions

Image input is provided inside messages[].content:

Parameter	Required	Description
model	Yes	A vision-capable model name, for example qwen3.6-35b-a3b or gpt-4o-mini.
messages[].content[]	Yes	An array containing text and image_url content parts.
type: "text"	Yes	The text prompt or instruction.
type: "image_url"	Yes	The image input part.
image_url.url	Yes	Either an HTTPS image URL or a base64 data URL such as data:image/png;base64,....
image_url.detail	Recommended	low, high, or auto. Use low by default for cost control.

Choosing image detail

Detail	When to use	Cost impact
low	Classification, broad scene description, simple screenshots, routing, quick checks	Lowest image-token usage; recommended default.
high	OCR-like tasks, small text, dense UI screenshots, charts, detailed diagrams, fine-grained defects	Higher image-token usage and higher managed-model cost.
auto	When you want the provider to choose	Cost can vary; avoid for strict cost control.

Cost estimation

For self-hosted models (qwen3.6-35b-a3b and qwen3.6-35b-a3b-think), CAIP currently charges no additional per-token usage cost, but throughput is capacity-limited.

For managed models, image input is billed as input tokens for the selected model. A practical estimate for OpenAI-style vision pricing is:

Detail	Approximate image input tokens	Example with gpt-4o at ~$5 / 1M input tokens
low	~85 input tokens per image	85 × $5 / 1,000,000 = **$0.00043 per image** plus prompt and output tokens
high	Starts with the low-detail budget, then adds higher-resolution image tiles/crops	More expensive; use only when low detail is insufficient

Example estimate for a simple request with one image, detail: "low", a 100-token text prompt, and a 300-token response on gpt-4o:

• Input: (85 image tokens + 100 text tokens) × $5 / 1,000,000 ≈ $0.00093
• Output: 300 tokens × $15 / 1,000,000 ≈ $0.00450
• Total estimate: ~$0.00543

Pricing can differ

CAIP pricing may differ from public provider prices. Use the model catalogue and contact CAIP for contractual pricing. The estimate above is intended for order-of-magnitude planning.

Curl examples

Image by HTTPS URL

curl -X POST \
  'https://llm.api.caip.bmw.cloud/v1/chat/completions' \
  -H 'Authorization: Bearer {Your_apikey}' \
  -H 'Content-Type: application/json' \
  -d '{
  "model": "qwen3.6-35b-a3b",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "Describe the relevant objects in this image."},
        {
          "type": "image_url",
          "image_url": {
            "url": "https://example.com/image.jpg",
            "detail": "low"
          }
        }
      ]
    }
  ],
  "disable_fallbacks": true
}'

Local image as base64 data URL

IMAGE_BASE64=$(base64 -w 0 ./screenshot.png)

curl -X POST \
  'https://llm.api.caip.bmw.cloud/v1/chat/completions' \
  -H 'Authorization: Bearer {Your_apikey}' \
  -H 'Content-Type: application/json' \
  -d "{
  \"model\": \"gpt-4o-mini\",
  \"messages\": [
    {
      \"role\": \"user\",
      \"content\": [
        {\"type\": \"text\", \"text\": \"Summarize this screenshot. Use low detail unless small text is required.\"},
        {
          \"type\": \"image_url\",
          \"image_url\": {
            \"url\": \"data:image/png;base64,${IMAGE_BASE64}\",
            \"detail\": \"low\"
          }
        }
      ]
    }
  ]
}"

OpenAI SDK examples in Python

Image by HTTPS URL

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://llm.api.caip.bmw.cloud/v1",
    api_key=os.environ["LLM_API_KEY"],
)

response = client.chat.completions.create(
    model="qwen3.6-35b-a3b",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Describe the relevant objects in this image."},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://example.com/image.jpg",
                        "detail": "low",
                    },
                },
            ],
        }
    ],
    extra_body={"disable_fallbacks": True},
)

print(response.choices[0].message.content)
print("served_by:", response.model)

Local image as base64 data URL

import base64
import os
from pathlib import Path

from openai import OpenAI

client = OpenAI(
    base_url="https://llm.api.caip.bmw.cloud/v1",
    api_key=os.environ["LLM_API_KEY"],
)

image_bytes = Path("screenshot.png").read_bytes()
image_base64 = base64.b64encode(image_bytes).decode("utf-8")

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Summarize this screenshot."},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/png;base64,{image_base64}",
                        "detail": "low",
                    },
                },
            ],
        }
    ],
)

print(response.choices[0].message.content)
print("served_by:", response.model)

Best practices

• Prefer qwen3.6-35b-a3b when the self-hosted model quality is sufficient and you want zero additional usage cost.
• For managed models, start with gpt-4o-mini, gpt-41-mini, gpt-5-mini, or qwen3-vl-flash before using larger models.
• Always set detail: "low" first; switch to high only after verifying low detail is not enough.
• Resize very large images before sending them if fine detail is not required.
• Avoid sending unnecessary or sensitive image regions; crop the image to the relevant area when possible.
• Check response.model to confirm the request was served by the intended model.