Category: provider

Model Studio Qwen VL (Image Understanding)

Validation

mkdir -p output/aliyun-qwen-vl
python -m py_compile skills/ai/multimodal/aliyun-qwen-vl/scripts/analyze_image.py && echo "py_compile_ok" > output/aliyun-qwen-vl/validate.txt

Pass criteria: command exits 0 and output/aliyun-qwen-vl/validate.txt is generated.

Output And Evidence

• Save raw model responses and normalized extraction results to output/aliyun-qwen-vl/.
• Include input image reference and prompt for traceability.

Use Qwen VL models for image input + text output understanding tasks via DashScope compatible-mode API.

Prerequisites

• Install dependencies (recommended in a venv):

python3 -m venv .venv
. .venv/bin/activate
python -m pip install requests

• Set DASHSCOPE_API_KEY in environment, or add dashscope_api_key to ~/.alibabacloud/credentials.

Critical model names

Prefer the Qwen3 VL family:

• qwen3-vl-plus
• qwen3-vl-flash

When you need explicit "latest" routing or reproducible snapshots, use supported aliases/snapshots from the official model list, such as:

• qwen3-vl-plus-latest
• qwen3-vl-plus-2025-12-19
• qwen3-vl-flash-2026-01-22
• qwen3-vl-flash-latest

Legacy names still seen in some workloads:

• qwen-vl-max-latest
• qwen-vl-plus-latest

For OCR-specialized extraction, prefer skills/ai/multimodal/aliyun-qwen-ocr/ instead of using the general VL skill.

Normalized interface (multimodal.chat)

Request

• prompt (string, required): user question/instruction about image.
• image (string, required): HTTPS URL, local path, or data: URL.
• model (string, optional): default qwen3-vl-plus.
• max_tokens (int, optional): default 512.
• temperature (float, optional): default 0.2.
• detail (string, optional): auto/low/high, default auto.
• json_mode (bool, optional): return JSON-only response when possible.
• schema (object, optional): JSON Schema for structured extraction.
• max_retries (int, optional): retry count for 429/5xx, default 2.
• retry_backoff_s (float, optional): exponential backoff base seconds, default 1.5.

Response

• text (string): primary model answer.
• model (string): model actually used.
• usage (object): token usage if returned by backend.

Quickstart

python skills/ai/multimodal/aliyun-qwen-vl/scripts/analyze_image.py \
  --request '{"prompt":"Summarize the main content in this image","image":"https://example.com/demo.jpg"}' \
  --print-response

Using local image:

python skills/ai/multimodal/aliyun-qwen-vl/scripts/analyze_image.py \
  --request '{"prompt":"Extract key information from the image","image":"./samples/invoice.png","model":"qwen3-vl-plus"}' \
  --print-response

Structured extraction (JSON mode):

python skills/ai/multimodal/aliyun-qwen-vl/scripts/analyze_image.py \
  --request '{"prompt":"Extract fields: title, amount, date","image":"./samples/invoice.png"}' \
  --json-mode \
  --print-response

Structured extraction (JSON Schema):

python skills/ai/multimodal/aliyun-qwen-vl/scripts/analyze_image.py \
  --request '{"prompt":"Extract invoice fields","image":"./samples/invoice.png"}' \
  --schema skills/ai/multimodal/aliyun-qwen-vl/references/examples/invoice.schema.json \
  --print-response

cURL (compatible mode)

curl -sS https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model":"qwen3-vl-plus",
    "messages":[
      {
        "role":"user",
        "content":[
          {"type":"image_url","image_url":{"url":"https://example.com/demo.jpg"}},
          {"type":"text","text":"Describe this image and list executable actions"}
        ]
      }
    ],
    "max_tokens":512,
    "temperature":0.2
  }'

Output location

• If --output is set, JSON response is saved to that file.
• Default output dir convention: output/aliyun-qwen-vl/.

Smoke test

python tests/ai/multimodal/aliyun-qwen-vl-test/scripts/smoke_test_qwen_vl.py \
  --image ./tmp/vl_test_cat.png

Error handling

Operational guidance

• For stable production behavior, pin snapshot model IDs instead of pure -latest.
• Compress very large images before upload to reduce latency and cost.
• Add explicit extraction constraints in prompt (fields, JSON shape, language).
• For OCR-like output, ask for confidence notes and unresolved text markers.

Workflow

1. Confirm user intent, region, identifiers, and whether the operation is read-only or mutating.
2. Run one minimal read-only query first to verify connectivity and permissions.
3. Execute the target operation with explicit parameters and bounded scope.
4. Verify results and save output/evidence files.

References

• Source list: references/sources.md
• API notes: references/api_reference.md