Alibaba Cloud PAI

What is Alibaba Cloud PAI?

AliCloud’s AI platform, PAI – which stands for Platform for Artificial Intelligence – is a cloud-based platform for machine learning, deep learning, and large-model development, designed for developers and enterprises. It covers all aspects such as data preparation, development, training, fine-tuning, evaluation, deployment, inference optimization, asset management, and operational monitoring; it is ideal for turning AI experiments into manageable production services.

PAI is neither a single-model API nor a regular chat tool. The platform is composed of multiple modules: iTAG is responsible for data annotation, Designer offers a drag-and-drop interface for model creation, DSW provides Notebook and VS Code-based development environments, DLC handles distributed training, Model Gallery is used for managing and fine-tuning pre-trained models, and EAS is responsible for deploying models as online inference services.

Data annotation iTAG

iTAG supports the annotation of images, text, videos, audio, and multimodal data; it offers predefined templates, custom question components, task assignment, as well as quality control and management of annotation results. Companies can also request large-scale hosted annotation services.

The basic use of the annotation platform can be enabled for free; for paid annotation services on a scale of 100,000 or more, it is necessary to submit a request. Before uploading real business data, it is essential to carry out data masking, authorization, and access control – data containing personal information must not be given directly to unauthorized annotators.

Designer low-code modeling

PAI-Designer offers a visual, drag-and-drop interface for machine learning workflows, and it includes a wide range of components for data processing, feature engineering, statistical analysis, as well as traditional and deep learning techniques. Users can connect data sources, configure algorithms, run experiments, and deploy the resulting services.

Designer is suitable for teaching, analysis, and rapid validation, but low-code does not mean that there is no need to understand data. Issues such as dataset leakage, class imbalance, incorrect metrics, and improper data splitting can still result in models that appear accurate but are not suitable for practical use.

DSW interactive modeling

DSW stands for Data Science Workshop; it offers Jupyter Notebook, VS Code, images, GPUs, and persistent storage in the cloud, making it suitable for data exploration, model development, debugging, training, and inference testing. Users can choose from the official images or configure their own dependencies as well as mount storage solutions such as OSS and NAS.

Public resources DSW are typically charged based on the duration for which the instances are running. Once development is complete, the instances should be stopped, and important code and models should be saved in persistent storage.

Some free cloud storage services may be deleted after the instance has been inactive for an extended period of time.

DLC distributed training

DLC is used to carry out training tasks with PyTorch, TensorFlow, MPI, etc.; it supports multi-GPU setups on a single machine, multi-machine configurations with multiple GPUs, queues, resource quotas, task logging, fault tolerance, and flexible scheduling. It is suitable for tasks such as the pre-training of large models, fine-tuning, working with visual models, and for teams that need to conduct batch experiments.

The cost of distributed training is determined by the model type, the number of GPUs, the location, and the duration of execution. A failed task can also consume resources; therefore, it is advisable to first test scripts, images, data paths, and Checkpoints using small datasets and low-resolution settings before starting the training process.

Model Gallery model library

The Model Gallery brings together models from Qwen and DeepSeek, as well as pre-trained models in the fields of vision, language, and AIGC; it also provides preset configurations for deployment, fine-tuning, evaluation, and inference for some of these models. The product page currently offers options for deploying the Qwen3.5 series in various parameter sizes.

Users can initiate DLC training and EAS deployment with no code or with only minimal configuration. The models themselves may be subject to different open-source licenses or commercial terms; before they are put into use in production, it is necessary to verify the licensing rules regarding their weights, any derived models, and the purposes for which they will be used.

Large model fine-tuning

PAI enables pre-trained models to be adapted for tasks such as corporate terminology handling, structured data extraction, customer service, classification, and content generation, through methods like full fine-tuning and LoRA. Model Gallery allows users to select datasets, training parameters, and hyperparameters, and it facilitates the deployment or evaluation of models after training.

Fine-tuning is not suitable for addressing all issues related to knowledge updates. Policies, prices, and data that change frequently are usually better suited for RAG.

Consider making fine-tunes only when a consistent output format, terminology style, or domain expertise is required.

Model evaluation and distillation

The platform allows for comparing base models with fine-tuned models using evaluation data; it is also possible to use large models to generate or label training samples, so as to train smaller models through the distillation process. This approach can help reduce inference costs and latency.

Synthetic data requires sampling for manual review; it is not permissible to copy teacher models incorrectly and without restriction to student models. The evaluation set should encompass real-world distributions, edge cases, and security issues, and it must be isolated from the training data.

EAS model online service

EAS, or Elastic Algorithm Service, enables the deployment of trained or pre-trained models as real-time, asynchronous, or offline inference services. The platform offers shared gateways, high-speed VPC connections, auto-scaling capabilities, as well as version control and monitoring functions; it also supports the use of certain large language models through OpenAI-compatible interfaces.

The public resource EAS is billed based on the duration for which a service instance is running, from its activation to shutdown; charges can be incurred even in the absence of any requests. Once testing is complete, the service must be stopped or deleted – it is not sufficient to simply close the browser.

Serverless inference

EAS offers a Serverless version for SD WebUI and ComfyUI; the deployment is free of charge, and billing is based on the actual number of seconds the requests are processed. It is suitable for image generation tasks that involve intermittent access.

At present, Serverless has a limited scope of support, so it cannot be inferred that all deployments of large models operate on a pay-per-request basis. The specific models, regions, cold start requirements, and unit prices are determined according to the page used to create the service.

Inference Acceleration PAI-Blade

PAI-Blade enhances the throughput of model inference and reduces latency through graph optimization, operator fusion, quantization, and the integration of hardware and software; it is suitable for applications in vision, natural language processing, and large model services. PAI-ACC, on the other hand, accelerates data processing, training, communication, and inference.

Accelerating and quantization may alter accuracy or model behavior. Before going live, it is necessary to use a business evaluation set to compare the original model with the optimized one, while also recording details regarding the hardware, images, and optimization parameters.

AI assets and MLOps

PAI offers capabilities for managing assets such as workspaces, datasets, models, images, pipelines, tasks, and services, thereby assisting teams in keeping track of training inputs, versions, parameters, and deployment results. Organizations can use RAM roles and workspaces to isolate different teams from one another.

To achieve reproducible training, it is also necessary to fix the code version used for submission, the container image, the random seed, the data version, and the dependencies. Saving only the final weights is not sufficient to explain the origin of the model.

FeatureStore feature platform

FeatureStore is used for the unified management of offline and online features, aiming to reduce feature discrepancies during training and in the service phase; it is suitable for applications such as recommendation systems, risk management, advertising, and search. Features can be reused across different models, and version control is provided for them.

For real-time features, attention must be paid to latency, expiration policies, data lineage, and privacy permissions. Faulty features can affect multiple online models at the same time, so rollback and monitoring mechanisms are necessary.

LangStudio and large-model applications

PAI also offers LangStudio-related functionalities for the development of large-model applications; it enables the management of prompts, models, knowledge, and application workflows, as well as the management of resources through OpenAPI. Combined with training and deployment tools, it allows companies to carry out model development and application implementation within the same cloud environment.

If the requirement is simply to make direct calls to APIs of hosted models such as Tongyi, Alibaba Cloud BaiLian might be a better choice; whereas PAI is more suitable when custom training, GPU resources, model assets, and dedicated deployment are needed.

Computing power and storage resources

PAI can make use of public resources, dedicated computing resources, intelligent computing resources, as well as various cloud-native resources, covering CPU, NVIDIA GPUs, and certain types of heterogeneous computing power. Data is typically stored in services such as OSS, NAS, CPFS, and MaxCompute.

In addition to the PAI billing, costs may also arise from object storage, file storage, elastic public networks, VPCs, logging services, and big data products. When estimating costs, it is necessary to take the entire architecture into account.

Free activation and trial for new users

• PAI products and the default workspace can be activated for free, with charging occurring only when those resources are actually used.
• Current new users can apply for a free usage quota for the DSW750, with the quota being renewed on a monthly basis over 3 months at 250 units per month;
• The DLC offers a trial period of 100 CU·H;
• EAS offers a discount package worth 500 yuan;
• The trial credit can only be used to offset specified free models and their corresponding modules, and it is not transferable between DSW, DLC, and EAS.
• Once the quota is exhausted or the trial period comes to an end, the resources that are still in use will switch to a pay-as-you-go pricing model.

Price and version comparison

Price of DSW resource packages

For the current introductory version of DSW 100, the initial purchase price of the resource package is 59 yuan; it is valid for 12 months, after which the original price applies upon further purchases. This resource package is available only in specific regions and for the introductory model versions.

Other GPUs, CPUs, and proprietary resources are billed based on their actual specifications. The prices of high-end GPUs vary significantly depending on availability, so it is not appropriate to specify fixed prices for them in the tool catalog over an extended period.

DLC training price

DLC can make use of public resources and pay on a post-payment basis based on the duration of usage; it is also possible to pre-order standard or high-performance resources and submit tasks by using quotas. The cost of training depends on the specifications of the instances, the number of nodes, the region, and the duration of use.

The 100 CU·H quota for new users is applicable only to specified free-tier specifications. Training with multiple machines and cards, as well as storage, networking, and logging, may incur additional charges.

EAS prices and savings plans

Current EAS new users can receive a 500-yuan discount for trial use. The introductory savings plan offers an initial purchase price of 59 yuan, with a 200-yuan discount available for public resource usage; this discount is valid for 12 months.

Pay-as-you-go public resources can be billed based on specified machines or combinations of CPU and memory.

The official reference prices for these combined resources are 0.194 yuan per core-hour for CPUs, and 0.024 yuan per GB-hour for memory; prices vary depending on the model and location.

Shared gateways are free of charge, while dedicated gateways, additional system disks, and public network resources are charged separately.

Example of costs for deploying the Qwen model

For the quick start guide of Qwen3-0.6B in Model Gallery, A10 GPU instances are used by default; the official example costs around 10.5 yuan per hour. Billing begins shortly after deployment, and the service should be stopped or deleted when it is not in use.

This price is merely an example for specific regions and specifications; it does not represent the uniform price for all Qwen, DeepSeek, or other large models. The larger the model parameters, the more GPUs and video memory are typically required.

API and Python SDK

PAI offers OpenAPI, general Alibaba Cloud SDKs, and advanced Python SDKs. The official Python SDK supports Python 3.8 and later versions; it allows for submitting training tasks, registering models, deploying services, and performing predictions. Additionally, some of the large-model services can be used as OpenAI-compatible clients.

SDK keys should be used with accounts and environment variables that grant the minimum required permissions; the AccessKey of the root account must not be included in Notebooks, images, or public code. For production use, it is necessary to implement token rotation, VPC access controls, and service rate limiting.

GitHub and the open-source status

The PAI cloud platform, scheduling system, console, and hosting services are not fully open-source products. Alibaba Cloud has made the pai-python-sdk available under an open-source license on GitHub, using the Apache 2.0 license, and it also maintains a repository with example code named pai-examples.

The fact that the SDKs and examples are open source means that the client-side code can be viewed and modified, but it does not imply that the PAI server, the training acceleration engine, or cloud resources can be deployed privately at no cost. It is also necessary to check the licenses for each model available in the Model Gallery.

Alibaba Cloud PAI Usage Guide

Complete a basic task.

1. Install and configure Alibaba Cloud PAI, and create separate projects and run names for the experiments;
2. Record the hyperparameters, data version, code version, and random seed in the training script;
3. Initialize the task using data annotation with iTAG and connect to the appropriate cloud or on-premises service;
4. Use the Designer low-code modeling tool to continuously record losses, metrics, logs, and hardware status;
5. Run a small-scale training session to check whether the number of steps, units, curves, and media samples are correct;
6. After completion, save the model, configurations, results, and instructions for reproducing the experiment;

Create reusable professional workflows

1. Standardize the naming conventions for projects, experiments, metrics, and labels;
2. Include data annotation with iTAG, low-code modeling using Designer, and interactive modeling with DSW in the training template;
3. Establish version associations for datasets, code, environments, and models;
4. Use the same evaluation set and comparison methodology to avoid focusing solely on the smooth curves;
5. Restrict the scope of uploading sensitive samples, prompts, and model outputs;
6. Set up exception alerts, retention policies, and conduct manual review before rolling them out to the team;

Which users is it suitable for?

• Data scientists who need cloud-based Notebooks and GPU development environments;
• An algorithm team responsible for fine-tuning, distillation, and distributed training of large models;
• Enterprise developers who need to deploy models as stable APIs;
• Analysts who wish to use the low-code Designer to carry out analyses for machine learning experiments;
• Platform teams that require workspace, permissions, assets, and MLOps governance;
• Companies that are working on projects related to recommendations, search, visual technologies, AIGC, and industry-specific models.

Product advantages

• It covers the entire process, including annotation, development, training, evaluation, and deployment;
• DSW, DLC, and EAS are suited to different stages of a project;
• Model Gallery offers practical applications for models such as Qwen and DeepSeek;
• Supports public resources, proprietary resources, and large-scale intelligent computing;
• It offers reasoning acceleration, elastic scaling, and enterprise-level access control.
• OpenAPI, Python SDK, and official examples are provided.

Restrictions and Precautions

• The PAI module has numerous components and complex billing mechanisms, resulting in a higher learning cost compared to using the model’s API directly.
• Costs can arise from the GPU, CPU, memory, system drive, OSS, NAS, public network, gateway, and logs.
• Pay-as-you-go resources may continue to incur charges even when not in use, so it is essential to set up budget alerts and stop or delete such testing resources.
• Enterprise data and models must also comply with regional, permission, licensing, and security requirements.

Frequently Asked Questions

Is Alibaba Cloud PAI free?

The platform can be activated for free, but there are charges for using resources such as computing power, storage, and deployment. New users can request trial credits for DSW, DLC, and EAS.

Can PAI be used to train large models?

Yes. DLC supports distributed training across multiple machines and cards, while Model Gallery enables the fine-tuning, evaluation, and deployment of models such as Qwen and DeepSeek.

What is the difference between DSW and DLC?

DSW is geared toward interactive Notebooks and development/debugging, while DLC is aimed at scheduled, scalable formal training tasks.

Is there a charge even if EAS is not used?

For regular public resource services, charging applies from the moment the service is started until it is stopped, and fees are incurred even in the absence of any requests; only specified Serverless services are charged based on the actual time taken to process tasks.

Does PAI support the OpenAI format?

Some large language models deployed via EAS offer an OpenAI-compatible interface for invocation; the specific path and capabilities are as specified in the documentation for the deployed model.

Is Alibaba Cloud PAI open source?

The platform itself is not open source; the official Python SDK and example repositories are open source, but this does not equate to the server and cloud platform being open source.