Z AI (Zhipu)

Z.ai is an AI lab that offers, through an API platform, the GLM family of language models, vision models, image generation models and video generation models, and also publishes open weights on Hugging Face.

The developer documentation, at docs.z.ai, describes a chat endpoint compatible with the OpenAI format, at api.z.ai/api/paas/v4, and includes an example with the OpenAI SDK pointed at that address.

The documentation's catalog covers:

• GLM language models, with a main model and Flash and FlashX variants.
• Vision models, which take images, videos and files together with text.
• CogView-4 and GLM-Image, for generating images.
• CogVideoX-3, for generating video from text or an image.

The documentation also gathers guides on function calling, structured output, web search and streaming tool calls, and offers an official Python SDK.

On Hugging Face, the zai-org organization publishes weights for the GLM family, and the cards of earlier models carry an MIT license.

Z.ai is for developers who want to integrate language, vision and media generation models through an API compatible with the OpenAI format, and for teams that prefer to run open GLM-family weights on their own infrastructure.

Usage profiles that the official documentation supports:

• Teams already using the OpenAI SDK, because the documentation shows the same SDK with the base URL changed.
• Products that call functions or tools, with guides on function calling, structured output and web search.
• Applications that analyze images, videos or files, with the vision models and the multimodal input of the Flash and FlashX variants.
• Products that generate images or videos, with CogView-4, GLM-Image and CogVideoX-3.
• Infrastructure teams that download zai-org weights from Hugging Face and serve them with vLLM, SGLang, Transformers, KTransformers or xLLM, the tools the model cards list.

Practical prerequisites. For the API you need an account and a key. Video generation is asynchronous, so your code must check for the result later. For open weights you need GPUs and must check the license of the exact model.

The main features of Z.ai are the chat API compatible with the OpenAI format, the GLM language models, multimodal input, image and video generation, tool guides, and open weights.

What the official documentation lists:

• A chat API at api.z.ai/api/paas/v4, compatible with the OpenAI format.
• GLM language models, with Flash and FlashX variants. The page for those variants lists function calling and multimodal input of images, videos and files.
• Vision models, described with native multimodal tool calling.
• Image generation: CogView-4, which uses a bilingual GLM encoder, and GLM-Image, whose response is an image URL that your code must download.
• Video generation: CogVideoX-3, which accepts text or an image and start and end frames, through an asynchronous call.
• Guides on function calling, structured output, web search and streaming tool calls.
• An official Python SDK, with functions, vision, image and video.
• Weights on Hugging Face, with an MIT license on the cards of earlier models.

The documentation also has a migration guide between generations of the main model.

Z.ai is worth it for anyone who wants text, vision, image and video on the same API platform and the option to run open weights, and less suited to anyone who needs a single synchronous response format for everything.

Strengths:

• One platform covers language, vision, image generation and video generation, according to the documentation's catalog.
• The chat endpoint is compatible with the OpenAI format, and the documentation shows the OpenAI SDK working with a changed base URL.
• The documentation has guides on function calling, structured output, web search and streaming tool calls.
• There are open weights on Hugging Face, with an MIT license on the cards of earlier models.

Points to watch:

• Video generation is asynchronous, so your code must check for the result later.
• The GLM-Image response is a URL, and your code must download the file.
• The license of the most recent models was not confirmed, and you need to read each model's card.
• The main model's name changes from generation to generation, and the documentation keeps a migration guide to follow along.

The alternatives to Z.ai are other labs with a language model API, some also with vision and media.

• Qwen: an Alibaba Cloud family with open weights, vision, audio and image models, and an API on Model Studio compatible with the OpenAI format.
• DeepSeek: an API in the OpenAI and Anthropic formats, with weights on Hugging Face and a context window of about 1 million tokens.
• Mistral AI: an API in the OpenAI format, with code, OCR and audio models, and open weights for part of the catalog.
• Cohere: an enterprise focus, with Embed and Rerank for search and VPC or on-premises deployment.
• OpenAI: its own models through an API, including image, with no downloadable weights for the main models.

The choice depends on which modality matters most. If the need is generating video through an API, Z.ai lists CogVideoX-3 in its documentation, and you need to check what each competitor offers on that point. If the criterion is weight licensing, compare model by model.