What is Alibaba Cloud Platform for AI (PAI)? - Platform For AI

Platform overview

Platform for AI (PAI) is an end-to-end AI development platform from Alibaba Cloud. PAI covers the full AI development lifecycle: model development, training, and deployment. PAI includes the following core modules:

Module	Description	Use case	Quick start
Model Gallery	Wraps EAS and DLC to let you train and deploy open-source large language models (LLMs) without writing code.	Zero-code training and deployment of open-source models	Model Gallery quick start
Elastic Algorithm Service (EAS)	Deploys trained models as online inference services with minimal configuration.	Model deployment as service endpoints	EAS quick start
Data Science Workshop (DSW)	Provides a cloud-based IDE (development machine). Users familiar with Notebook or VS Code can start developing models right away.	AI model development	DSW quick start
Deep Learning Containers (DLC)	Quickly creates distributed or single-node training jobs without manual machine provisioning or environment setup. Works with the same training scripts you run locally.	Distributed model training	DLC quick start
Machine Learning Designer	Includes 140+ built-in algorithm components for low-code visual modeling. Build models by dragging and dropping components.	Big data and AI model development	Designer quick start

Click a module name to learn more about its features.

Benefits

Full AI development lifecycle

Covers data annotation, model development, training, optimization, deployment, and AI operations management in a single platform.
Includes 140+ optimized built-in algorithm components.
Offers multiple development modes, deep integration with big data engines, multi-framework compatibility, and custom images.

Multiple open-source frameworks

Supports the Flink stream processing framework.
Provides the deep learning frameworks TensorFlow, PyTorch, Megatron, and DeepSpeed, deeply optimized from their open-source releases.
Works with mainstream open-source frameworks such as Spark, PySpark, and MapReduce.

Industry-leading AI optimization

Provides a high-performance sparse training framework that supports billions to tens of billions of sparse features and tens of billions to hundreds of billions of samples. The framework also supports distributed incremental training across 1,000+ workers.
Accelerates inference for models built on mainstream frameworks. PAI-Blade increases the acceleration ratio for models such as ResNet50 and Transformer+LM.

Flexible deployment options

Supports fully managed and semi-managed modes on the public cloud.
Offers AI high-performance computing clusters and lightweight deployment options.
Supports periodic scheduling through DataWorks, with separate production and development environments for data isolation.

Billing

Billing method	Description	Use case	Applicable modules
Pay-as-you-go	Pay after use, based on actual consumption.	Short-term or unpredictable workloads such as test environments, demand spikes, or early-stage projects.	DSW, DLC, EAS, Designer
Subscription	Prepaid monthly or yearly plans.	Long-term, steady workloads. Prepaying for a fixed period, such as one month or one year, gives you a lower unit price than with pay-as-you-go.	DSW, DLC, EAS
Resource plan	Prepaid quota package for a specific resource.	Heavy use of one resource type. Quota packages offer volume discounts.	DSW
Savings plan	Prepaid plan with a committed spend.	A fixed spending commitment over a set period, in exchange for a lower pay-as-you-go rate.	DSW, EAS
Duration-based billing (Serverless)	Pay after use. Charges apply only for the time the service actively processes requests. Deployment is free, and the service scales automatically based on request volume.	Variable request volume with Serverless deployment. Handles high concurrency and dynamic loads.	EAS

For more information, see Billing of AI computing resources.

New user guide

If you are new to PAI, start with Get started with PAI.

Use cases

LLM deployment and fine-tuning

AIGC

Retrieval-augmented generation (RAG)

AI agents

FAQ

Q: What do I do if a DSW instance fails to start or stop, and how do I release it?

For more information, see DSW FAQ - Instance start and release.

Q: Why do EAS service calls fail?

For more information, see EAS FAQ - Service invocation.