VLM

Create a ZAI instance with ZAI.create(), then call zai.chat.completions.createVision with a message containing a text part and an image_url part. The response content is available at response.choices[0].message.content.

Read the file with fs.readFileSync, convert the buffer to base64, and embed it as a data URL like data:image/png;base64,... in the image_url field. Base64 encoding is recommended over URLs for better performance and reliability.

No, the z-ai-web-dev-sdk must be used in backend code only. Importing it in client-side code will fail and risks exposing credentials, so always call it from server-side endpoints such as an Express.js API route.

The CLI supports PNG, JPEG, GIF, WebP, and BMP formats. The SDK also accepts video files via video_url and document files like PDF and DOCX via file_url, and multiple content types can be combined in one message.

Use the z-ai vision CLI for quick one-off image descriptions, testing, and simple automation. Use the SDK for multi-turn conversations, dynamic analysis in applications, batch processing, and production workflows requiring custom logic.

Poor results usually come from low-quality images or vague prompts. Use high-resolution images, be specific about what information you need, and structure complex requests with numbered lists or context about the image type.