Flink AI service (built-in models) - Realtime Compute for Apache Flink

Real-Time Compute for Apache Flink provides built-in AI models that you can call directly in Flink SQL jobs and Flink Agent without configuring an API key or endpoint.

Important

This feature is in whitelisted Beta. You must submit an application and be approved before you can use it. No separate activation is required. During the Beta, model invocations are free of charge. The billing methods and service models described below take effect after commercial release. For the release date, see product announcements.

Overview

Flink AI service (built-in models) is a managed AI model invocation capability in Real-Time Compute for Apache Flink. You can call built-in models in Flink SQL jobs and Flink Agent without configuring an API key, enabling streaming AI inference and embedding.

Core advantages

Feature	Description
Out-of-the-box	No need to configure an API-Key, endpoint, or PrivateLink.
Billed by token	Token-based billing with clear, controllable costs.
Multi-model support	Supports popular large models for text generation, visual understanding, translation, and embedding.
Global region coverage	Available in multiple regions inside and outside the Chinese mainland, with automatic cross-region access.

How it works

Built-in model vs. custom model (BYOK)

Dimension	Item	Built-in model (Recommended)	Custom model (Not recommended)
Access configuration	endpoint	Not required. The system configures it automatically.	Required.
	api-key	Not required. The system manages it.	You must provide and maintain it.
	Supported models	Only models in the built-in model list.	Any model compatible with the OpenAI/DashScope protocol.
	Prerequisites	The primary account must enable the Flink AI feature in the desired region.	You must enable the model service and obtain an API-Key.
Network and security	Access path	Access is direct over the Alibaba Cloud internal network. Traffic does not leave the VPC.	Requires public internet access.
	Data security	High security. Traffic is transmitted end-to-end over the Alibaba Cloud internal network.	Data is transmitted over the public internet. You are responsible for assessing the risks.
	Network configuration cost	Zero configuration.	Public internet access requires a NAT Gateway, an EIP, or a public IP address.
Performance	Invocation latency	Low latency. Direct internal connection avoids extra network hops.	Higher and less stable latency due to public network fluctuations.
Performance	Availability	SLA-backed with stable network and lower latency.	Depends on the model service provider's SLA and network link stability.
Billing	Model invocation fees	Billed by token.	Billed directly by the model service provider.
	Network fees	No extra fees.	You must pay for the NAT Gateway, EIP, and public network traffic.
	Billing consolidation	Consolidated in your Flink bill.	Model and network fees are billed separately.
Capability boundaries	Token limits	Up to 8K for input, and up to 2K for output.	Determined by the model service provider.

Enable and disable the service

Important

This feature is in whitelisted Beta. You must submit an application and be approved before you can use it. No separate activation is required.

Prerequisites

A primary account must enable the feature.
Once enabled in a region, all workspaces in that region can use it.
This feature is pay-as-you-go. You are charged only for model invocations, not for enabling the feature. For details, see

Log on to the Real-Time Compute for Apache Flink console. On the Flink AI Service page, click Enable Now.

Disable the service

On the Flink AI Service page, click Disable Service and confirm the following:

After you disable the feature, all jobs in the region can no longer use the built-in model service.
Before you disable the feature, make sure that no jobs depend on the built-in models.

Supported regions

Flink AI service region	Region ID	Endpoint region	Inference execution region
China (Beijing)	cn-beijing	China (Beijing)	Chinese mainland (dynamic scheduling)
China (Zhangjiakou)	cn-zhangjiakou
China (Ulanqab)	cn-wulanchabu
China (Shanghai)	cn-shanghai
China (Hangzhou)	cn-hangzhou
China (Shenzhen)	cn-shenzhen
China (Chengdu)	cn-chengdu
Singapore	ap-southeast-1	Singapore	Regions outside the Chinese mainland (dynamic scheduling)
China (Hong Kong)	cn-hongkong
Malaysia (Kuala Lumpur)	ap-southeast-3
Indonesia (Jakarta)	ap-southeast-5
Japan (Tokyo)	ap-northeast-1
UK (London)	eu-west-1
Germany (Frankfurt)	eu-central-1
UAE (Dubai)	me-east-1
Mexico	na-south-1

Note

Choose a region based on your data compliance requirements. The selected region determines where data is accessed and where inference runs.

Endpoint region: Determines the access point and data storage location.
Inference execution region: Where inference runs. Dynamically scheduled within a defined scope.

Built-in models

Inference models (chat/completions)

Model name	Use case	Input	Output	Supported regions
qwen3.6-plus	Visual understanding and text generation. The latest flagship Qwen model.	Text/Image/Video	Text	All
qwen3.6-flash	Visual understanding and text generation. A cost-effective model.	Text/Image/Video	Text	All
qwen3.5-plus	Visual understanding and text generation. A high-performance model.	Text/Image/Video	Text	All
qwen3.5-flash	Visual understanding and text generation. A fast, low-cost model.	Text/Image/Video	Text	All

Embedding models (embeddings)

Model name	Use case	Input	Output	Supported regions
text-embedding-v4	Text embedding	Text	Vector	All
qwen3-vl-embedding	Multimodal embedding	Image/Text/Video	Vector	Chinese mainland

Usage

Basic syntax

To use a built-in model, specify the task and model parameters in the CREATE MODEL statement. Do not specify endpoint or api-key.

CREATE MODEL model_name
INPUT (column_name STRING)
OUTPUT (column_name {STRING | ARRAY<FLOAT>})
WITH (
  'provider' = 'openai-compact',
  'task' = 'chat/completions | embeddings',
  'model' = '<model-name>'
);

For more details, see Model configuration.

Relationship with custom models

Scenario	endpoint	api-key	Behavior
Use a built-in model	Not specified	Not specified	Automatically uses the Flink-managed model service.
Use a custom model	Specified	Specified	Uses a custom model service (BYOK mode).

Your primary account must first enable the Flink AI feature before you can use built-in models. Otherwise, jobs that omit endpoint and api-key fail.

Limitations

Item	Limit	Description
Maximum input tokens	8K	The system default limit. If you specify `max-context-size`, the smaller value applies.
Maximum output tokens	2K	The system default limit. If you specify `max-tokens`, the smaller value applies.
Model scope	Built-in model list	Only models in the list are supported. Specifying any other model causes an error.

Use cases

Scenario	Recommended model	Description
Real-time text classification and tagging	qwen3.5-flash / qwen3.6-flash	Real-time sentiment analysis and classification tagging on streaming data.
Real-time information extraction	qwen3.5-plus / qwen3.6-plus	Extract structured information from unstructured text.
Real-time embedding	text-embedding-v4	Generate embeddings from streaming data for similarity searches.
Multimodal embedding	qwen3-vl-embedding	Generate unified embeddings for images, text, and videos.