Agent setup: If your agent doesn't auto-load skills (e.g. Claude Code),
see agent-compatibility.md once per session.

Qwen Vision (Image & Video Understanding)

Analyze images and videos using Qwen VL and QVQ models.
This skill is part of qwencloud/qwencloud-ai.

Skill directory

Use this skill's internal files to execute and learn. Load reference files on demand when the default path fails or you need details.

Location	Purpose
INLINECODE0	Image/video understanding, multi-image, thinking mode
INLINECODE1

Visual reasoning (QVQ, chain-of-thought, streaming) |
| scripts/ocr.py | OCR text extraction |
| scripts/vision_lib.py | Shared helpers (base64, upload, streaming) |
| references/execution-guide.md | Fallback: curl, code generation |
| references/curl-examples.md | Curl for base64, multi-image, video, OCR |
| references/visual-reasoning.md | QVQ and thinking mode details |
| references/prompt-guide.md | Query prompt templates by task, thinking mode decision |
| references/ocr.md | OCR parameters and examples |
| references/sources.md | Official documentation URLs |
| references/agent-compatibility.md | Agent self-check: register skills in project config for agents that don't auto-load |

Security

NEVER output any API key or credential in plaintext. Always use variable references ($DASHSCOPE_API_KEY in shell, os.environ["DASHSCOPE_API_KEY"] in Python). Any check or detection of credentials must be non-plaintext: report only status (e.g. "set" / "not set", "valid" / "invalid"), never the value. Never display contents of .env or config files that may contain secrets.

When the API key is not configured, NEVER ask the user to provide it directly. Instead, help create a .env file with a placeholder (DASHSCOPE_API_KEY=sk-your-key-here) and instruct the user to replace it with their actual key from the QwenCloud Console. Only write the actual key value if the user explicitly requests it.

Key Compatibility

Scripts require a standard QwenCloud API key (sk-...). Coding Plan keys (sk-sp-...) cannot be used for direct API calls and do not support dedicated vision models (qwen3-vl-plus, qvq-max, etc.). The scripts detect sk-sp- keys at startup and print a warning. If qwencloud-ops-auth is installed, see its references/codingplan.md for full details.

Model Selection

Model	Use Case
qwen3.5-plus	Preferred — unified multimodal (text+image+video). Thinking on by default.
qwen3.5-flash

Fast multimodal — cheaper, faster. Thinking on by default. | | qwen3-vl-plus | High-precision — object localization (2D/3D), document/webpage parsing. | | qwen3-vl-flash | Fast vision — lower latency, 33 languages. | | qvq-max | Visual reasoning — chain-of-thought for math, charts. Streaming only. | | qwen-vl-ocr | OCR — text extraction, table parsing, document scanning. | | qwen-vl-max | Qwen2.5-VL — best-performing in 2.5 series. | | qwen-vl-plus | Qwen2.5-VL — faster, good balance of performance and cost, 11 languages. |

1. 1. User specified a model → use directly.
2. Consult the qwencloud-model-selector skill when model choice depends on requirement, scenario, or pricing.
3. No signal, clear task → qwen3.5-plus. Use qwen3-vl-plus for precise localization or 3D detection.

⚠️ Important: The model list above is a point-in-time snapshot and may be outdated. Model availability
changes frequently. Always check the official model list
for the authoritative, up-to-date catalog before making model decisions.

Execution

Prerequisites

• - API Key: Check that DASHSCOPE_API_KEY (or QWEN_API_KEY) is set using a non-plaintext check only (e.g. in shell:

[ -n "$DASHSCOPE_API_KEY" ]; report only "set" or "not set", never the key value). If not set: run the * qwencloud-ops-auth* skill if available; otherwise guide the user to obtain a key from QwenCloud Console and set it via .env file ( echo 'DASHSCOPE_API_KEY=sk-your-key-here' >> .env in project root or current directory) or environment variable. The script searches for .env in the current working directory and the project root. Skills may be installed independently — do not assume qwencloud-ops-auth is present.

• - Python 3.9+ (stdlib only, no pip install needed)

Environment Check

Before first execution, verify Python is available:

CODEBLOCK0

If python3 is not found, try python --version or py -3 --version. If Python is unavailable or below 3.9, skip to Path 2 (curl) in execution-guide.md.

Default: Run Script

Script path: Scripts are in the scripts/ subdirectory of this skill's directory (the directory containing this SKILL.md). You MUST first locate this skill's installation directory, then ALWAYS use the full absolute path to execute scripts. Do NOT assume scripts are in the current working directory. Do NOT use cd to switch directories before execution. Shared infrastructure lives in scripts/vision_lib.py.

Execution note: Run all scripts in the foreground — wait for stdout; do not background.

Discovery: Run python3 <this-skill-dir>/scripts/analyze.py --help (or reason.py, ocr.py) first to see all available arguments.

Script	Purpose	Default Model
INLINECODE37	Image understanding, multi-image, video, thinking mode, high-res	INLINECODE38
INLINECODE39

Visual reasoning with chain-of-thought, video reasoning (always streaming) | qvq-max |
| scripts/ocr.py | OCR text extraction from documents, receipts, tables | qwen-vl-ocr |

Input type fields (use exactly one in --request JSON):

Field	Use for	Example
INLINECODE44	Single image (URL or local path)	INLINECODE45
INLINECODE46

Multi-image comparison (array) | "images": ["a.jpg", "b.jpg"] |
| "video" | Video file (URL or local path) | "video": "clip.mp4" |
| "video_frames" | Video as frame array | "video_frames": ["f1.jpg", "f2.jpg"] |

⚠️ Common mistake: Do NOT use "image" for video files — use "video" instead.

CODEBLOCK1

Argument	Description
INLINECODE54	JSON request body
INLINECODE55

Load request from file |
| --output path | Save response JSON to path |
| --print-response | Print response to stdout |
| --stream | Enable streaming (auto for thinking/QVQ) |
| --upload-files | Upload local files to temp storage (for files > 7 MB) |
| --schema path.json | JSON Schema for structured extraction |

Verify Result

• - Exit code 0 + output contains valid JSON with choices field → success
• Non-zero exit, HTTP error, empty response, or JSON with "code"/"message" error → fail
• Post-execution check: When --output is used, verify the response JSON file exists and contains expected content
• MANDATORY — stderr signal check: After confirming the result, scan the command's stderr output for [ACTION_REQUIRED] or [UPDATE_AVAILABLE]. If either signal is present, you MUST follow the instructions in Update Check below before responding to the user.

On Failure

If scripts fail, match the error output against the diagnostic table below to determine the resolution. If no match, read execution-guide.md for alternative paths: curl commands (Path 2), code generation (Path 3), and autonomous resolution (Path 5).

If Python is not available at all → skip directly to Path 2 (curl) in execution-guide.md.

Error Pattern	Diagnosis	Resolution
INLINECODE68	Python not on PATH	Try python or py -3; install Python 3.9+ if missing
INLINECODE71

Script version check failed | Upgrade Python to 3.9+ |
| SyntaxError near type hints | Python < 3.9 | Upgrade Python to 3.9+ |
| QWEN_API_KEY/DASHSCOPE_API_KEY not found | Missing API key | Obtain key from QwenCloud Console; add to .env: echo 'DASHSCOPE_API_KEY=sk-...' >> .env; or run qwencloud-ops-auth if available |
| HTTP 401 | Invalid or mismatched key | Run qwencloud-ops-auth (non-plaintext check only); verify key is valid |
| SSL: CERTIFICATE_VERIFY_FAILED | SSL cert issue (proxy/corporate) | macOS: run Install Certificates.command; else set SSL_CERT_FILE env var |
| URLError / ConnectionError | Network unreachable | Check internet; set HTTPS_PROXY if behind proxy |
| HTTP 429 | Rate limited | Wait and retry with backoff |
| HTTP 5xx | Server error | Retry with backoff |
| PermissionError | Can't write output | Use --output to specify writable directory |

File Input

The API accepts: HTTP/HTTPS URL, Base64 data URI, and oss:// URL. Local file paths are NOT directly supported — scripts handle conversion automatically. Pass local paths directly; no manual upload step needed.

Large file rule: If the local file is >= 7 MB, always add --upload-files. Base64 encoding inflates size by ~33% and will exceed the 10 MB API limit. Small files (including short video clips < 7 MB) can use the default base64 path.

Method	When to use	How
Online URL	File already hosted	Pass URL directly — preferred for large files
Base64 (default)

Local files < 7 MB (images or short video clips) | Script auto-converts to data: URI |
| Temp upload | Local files >= 7 MB | Add --upload-files flag → uploads to DashScope temp storage (oss:// URL, 48h TTL) |

Production: Default temp storage has 48h TTL and 100 QPS upload limit — not suitable for production, high-concurrency, or load-testing. To use your own OSS bucket, set QWEN_TMP_OSS_BUCKET and QWEN_TMP_OSS_REGION in .env, install pip install alibabacloud-oss-v2, and provide credentials via QWEN_TMP_OSS_AK_ID / QWEN_TMP_OSS_AK_SECRET or the standard OSS_ACCESS_KEY_ID / OSS_ACCESS_KEY_SECRET. Use a RAM user with least-privilege (oss:PutObject + oss:GetObject on target bucket only). The --upload-files flag is still required for vision scripts to trigger upload. If qwencloud-ops-auth is installed, see its references/custom-oss.md for the full setup guide.

Input from Other Skills

When the input file comes from another skill's output (e.g., image-gen, video-gen):

• - Pass the URL directly (e.g., "image": "<image_url from image-gen>") — do NOT download the URL first
• Downloading and re-passing as a local path wastes bandwidth and triggers unnecessary base64 encoding or OSS upload
• All URL types are supported: https://, oss://, INLINECODE107

Thinking Mode

Model	Thinking Default	Notes
INLINECODE108 / INLINECODE109	On	Disable with enable_thinking: false for simple tasks.
INLINECODE111 / INLINECODE112

Off | Enable with enable_thinking: true. | | qvq-max | Always on | Streaming output required. |

See visual-reasoning.md for details.

OCR (qwen-vl-ocr)

Optimized for text extraction. Supports multi-language, skewed images, tables, formulas. See ocr.md for parameters and examples.

Input Limits

Images: BMP/JPEG/PNG/TIFF/WEBP/HEIC. Min 10px sides, aspect ratio <= 200:1. Max 20 MB (URL, Qwen3.5) / 10 MB (others).

Videos: MP4/AVI/MKV/MOV/FLV/WMV. Duration 2s–2h (Qwen3.5) / 2s–10min (others). Max 2 GB (URL) / 10 MB (base64). fps range [0.1, 10], default 2.0.

Error Handling

HTTP	Meaning	Action
401	Invalid or missing API key	Run qwencloud-ops-auth if available; else prompt user to set key (non-plaintext check only)
400

Bad request (invalid format) | Verify messages format and image URL/format | | 429 | Rate limited | Retry with exponential backoff | | 5xx | Server error | Retry with exponential backoff |

Output Location

Prefer the current working directory. Default subdirectory: ./output/qwencloud-vision/.

Write prohibition: NEVER write output files into this skill's installation directory or any skills/ hierarchy. All generated content must go to output/ under the current working directory or a user-specified path.

Update Check (MANDATORY Post-Execution)

After script execution, check stderr for action signals. The script automatically detects update status and emits structured signals. You MUST act on them. This section also applies when the user explicitly asks to check for updates.

Signal: [ACTION_REQUIRED] qwencloud-update-check skill is not installed

When stderr contains this signal, you MUST present the following options to the user and wait for their response before proceeding. Do NOT silently skip this.

The update-check skill is not installed. It keeps you informed about new qwencloud/qwencloud-ai features and bug fixes.
1) Install now — I'll run the install command for you
2) Skip for now — do nothing this time
3) Never remind me — permanently suppress this prompt

After the user responds:

• - User chooses 1 → run: npx skills add QwenCloud/qwencloud-ai --skill qwencloud-update-check -y, then run: INLINECODE120
• User chooses 2 → run: python3 <this-skill-scripts-dir>/gossamer.py --dismiss, then continue normally
• User chooses 3 → run: python3 <this-skill-scripts-dir>/gossamer.py --never-install, then continue normally

Signal: [UPDATE_AVAILABLE]

When stderr contains this signal, you MUST append a brief update notice to your response, including the version info and the update command shown in the stderr output.

No signal in stderr

If stderr contains neither [ACTION_REQUIRED] nor [UPDATE_AVAILABLE], no action is needed — the skill is installed and up to date (or cached within 24h).

Explicit user request

When the user explicitly asks to check for updates (e.g. "check for updates", "check version"):

1. 1. Look for qwencloud-update-check/SKILL.md in sibling skill directories.
2. If found — run: python3 <qwencloud-update-check-dir>/scripts/check_update.py --print-response and report the result.
3. If not found — present the install options above.

References

• - execution-guide.md — Fallback paths (curl, code generation, autonomous)
• curl-examples.md — Curl templates (base64, multi-image, video, OCR)
• api-guide.md — API supplementary guide
• visual-reasoning.md — QVQ visual reasoning guide
• ocr.md — Qwen-VL-OCR text extraction guide
• sources.md — Official documentation URLs