2025
October
Release date | Feature | Description | References |
2025-10-17 | Lingjun AI Computing Service GU7 series now supports the 570 driver. | The Lingjun AI Computing Service GU7 series now supports the 570 driver. The PAI training service for the GU7 series supports multiple drivers, such as 530, 550, and 570. Select different drivers when submitting a task without reinstallation. This feature fully utilizes the capabilities of the Serverless platform to provide an optimal user experience. | |
2025-10-17 | Scenario-based deployment of Dify on PAI-EAS is released. | Dify is an open source development platform for large language model (LLM) applications. It helps developers, enterprises, and non-technical personnel quickly build, deploy, and manage applications based on generative AI. Deploy the open source version of the Dify platform on EAS with a single click. This supports using the WebUI and related API calls to quickly build, deploy, and manage generative AI-based applications. | |
2025-10-17 | Message notifications are supported for events throughout lifecycle of DSW instances. | 1. Workspace administrators can configure message notification rules in Workspace Configuration > Event Notification. Select instance statuses, image saving statuses, and notification targets such as DingTalk, text message, phone call, or WeCom. 2. When a status event in a notification rule occurs, you receive real-time message notifications to take prompt action. | |
2025-10-17 | The model distillation feature is released. | This feature uses the PAI self-developed EasyDistill algorithm library to provide productized model distillation. It transfers the capabilities of a large teacher model to a small student model through distillation. It also supports online synthesis of distillation data, including instruction data for non-inference models and chain-of-thought data for inference models. This helps the small model approach or match the performance of the large model on specific tasks, improving model performance and reducing deployment costs. |
September
Release date | Feature | Description | References |
2025-09-19 | ArtLab releases the Design Agent. | Easily complete high-quality image generation, video production, and fine-grained image editing using natural language instructions. This significantly lowers the barrier to creative expression. Unleash the creative power of natural language and redefine the AIGC design workflow. | |
2025-09-19 | EAS computing power detection and fault tolerance feature is released. | The EAS computing power detection and fault tolerance feature comprehensively inspects resources involved in inference. It automatically isolates faulty nodes and triggers background automated O&M processes. This effectively reduces the likelihood of encountering issues during the initial phase of service inference and improves the success rate of inference deployment. | |
2025-09-17 | Data development now supports direct execution of PAI Flow nodes. | PAI Flow provides end-to-end machine learning workflow development capabilities. It offers the same workflow features as the visual modeling Designer of Platform for AI (PAI) and can periodically schedule workflows. |
August
Release date | Feature | Description | References |
2025-08-25 | EAS supports expert parallel (EP) deployment. | EAS now implements expert parallel (EP) deployment for Mixture-of-Experts (MoE) models such as DeepSeek-R1. It supports inference engines like vLLM and SGLang, enabling customers to overcome hardware limitations, improve resource utilization, and increase system throughput. | Deploy MoE models using expert parallelism and Prefill-Decode separation |
2025-08-15 | Model Evaluation Center v1.0 is released. | Use this out-of-the-box feature to complete the end-to-end model evaluation pipeline without code development. Quickly assess if a model's capabilities are suitable for your business scenarios. | |
2025-08-13 | AI resource groups (Lingjun AI Computing Service) now support pay-as-you-go and savings plans. | AI resource groups (Lingjun AI Computing Service) now support pay-as-you-go purchases. When combined with savings plans, this feature automatically applies discounts based on the subscription duration. The longer the subscription, such as 1, 3, or 5 years, the greater the discount. This provides users with a more flexible and cost-effective way to use resources. | |
2025-08-13 | DataJuicer on DLC is officially released. | DLC supports submitting DataJuicer framework tasks. It efficiently performs large-scale data cleaning, filtering, transformation, and enhancement using multiple operators (100+), multiple scales (single-node and multi-node), and high availability (self-healing). This enables text and multi-modal data processing capabilities for large model scenarios. | |
2025-08-11 | DLC self-developed Custom task (v1.0) is officially released. | The DLC self-developed distributed framework, Custom, supports PAI scheduling policies and self-healing capabilities. It also provides advanced features such as custom roles, success policies, and extended ports. This meets the computing demands of various business scenarios, including post-training for large models and autonomous driving. | |
2025-08-08 | The model weight service is released. | The model weight service significantly reduces cold start and scale-out times. It addresses the industry-wide challenge of long model loading times and overcomes the performance bottleneck of ultra-large-scale LLM deployment. | |
2025-08-07 | EAS releases the Prefill-Decode separation feature. | EAS releases the Prefill-Decode (PD) separation feature. It includes multiple deployment modes such as static PD separation and dynamic PD separation. It supports various inference engines like vLLM, SGLang, and BladeLLM to help customers reduce inference latency. |
July
Release date | Feature | Description | References |
2025-07-10 | DSW supports distributed development and debugging environments. | This feature helps you debug and verify distributed tasks to create a more efficient development and training workflow.
|
June
Release date | Feature | Description | References |
2025-06-10 | ArtLab supports building and sharing AIGC applications based on ComfyUI. | PAI-ArtLab upgrades its enterprise-level AIGC application capabilities. 1. Supports custom building and publishing of AIGC applications based on ComfyUI workflows. 2. Share AIGC applications from the ArtLab platform as out-of-the-box AIGC applications for PC and H5 mobile clients. | |
2025-06-05 | Data development supports Platform for AI (PAI) Flow. | This feature unifies the entry points for big data development and AI products. It enhances the deep integration between PAI Flow and big data engines to achieve integrated big data and AI development.
|
May
Release date | Feature | Description | References |
2025-05-30 | QuickStart > Model Gallery is launched in the Malaysia (Kuala Lumpur) region. | Model Gallery is launched in the Malaysia (Kuala Lumpur) region. Model Gallery integrates high-quality pre-trained models from various open source AI communities. It helps you get started quickly and use PAI for model training and deployment. |
April
Release date | Feature | Description | References |
2025-04-04 | Image management now supports building custom images. | PAI officially launches the custom image building feature. Developers and enterprise customers can flexibly install dependencies based on existing images or build custom images using a custom Dockerfile. The images are automatically pushed to ACR and registered on the PAI platform. This meets personalized customization needs and eliminates the tedious process of local building and uploading. | |
2025-04-01 | DSW is officially launched in US (Virginia) and US (Silicon Valley). | DSW is now available in the US (Virginia) and US (Silicon Valley) regions. You can use it on the console as needed. |
March
Release date | Feature | Description | References |
2025-03-28 | LangStudio 1.0 is officially released. | This version adds the following features based on LangStudio 0.1: 1. Knowledge base creation and management: Create and sync knowledge bases in the console and use them when building application flows. 2. Use PAI-DSW as the application development environment: Provides DSW-based Notebook and WebIDE environments for application flow development. 3. Evaluate application flow performance: Pre-built evaluation templates support offline performance evaluation and online service performance evaluation for applications. 4. Deployed applications support conversation history: Use a cloud database or local storage for conversation history, and manage and export the history. | |
2025-03-28 | DSW supports dynamic mounting of NAS datasets and storage paths. | DSW supports dynamic storage mounting. You can mount or unmount NAS datasets without restarting the instance. | |
2025-03-28 | Deep Learning Containers (DLC) supports mounting OSS data sources using ossfs. | Deep Learning Containers (DLC) supports mounting OSS data sources using ossfs. This provides good OSS read and write performance for compute-intensive tasks such as autonomous driving, which typically involve sequential and random reads, and sequential append writes. | |
2025-03-27 | AI computing power node statuses are upgraded. | The statuses of computing power nodes are optimized. A new status code is added to prohibit scheduling, which improves your user experience. | |
2025-03-19 | When you submit Ray tasks in Deep Learning Containers (DLC), you can now use custom roles. | When you submit Ray framework tasks in Deep Learning Containers (DLC), you can customize the Worker role to enable hybrid execution of heterogeneous resources. | |
2025-03-19 | Resource quotas support scaling out and scaling in specified nodes. | Scaling for resource quotas now supports node-level operations. This makes computing power management, reallocation, and transfer operations between quotas more flexible. | |
2025-03-07 | PAI training service is officially launched in US (Silicon Valley). | Deep Learning Containers (DLC) and AI resource quotas are now available in the US (Silicon Valley) region. You can use resource quotas and public resources (pay-as-you-go) to submit training tasks. |
February
Release date | Feature | Description | References |
2025-02-28 | When you mount storage instances (such as NAS and CPFS) in Deep Learning Containers (DLC), you can now configure access control lists. | When you mount Alibaba Cloud storage instances such as OSS, NAS, and CPFS in Deep Learning Containers (DLC), you can configure access control lists. This supports fine-grained permission management for your storage instances. | |
2025-02-21 | AI scheduling engine v2.0 implements multi-level task preemption. | The PAI scheduling engine, based on resource quotas, uses task-level classification (such as training, inference, development, and priority) and a dynamic priority evaluation algorithm to trigger a preemption mechanism. This ensures that high-priority tasks can be executed quickly. Combined with AIMaster's preemptive rollback technology, interrupted tasks automatically save their intermediate state and enter a queue. They are then prioritized for resumption after resources are released. This achieves efficient scheduling in resource-constrained scenarios. | |
2025-02-10 | When you mount Alibaba Cloud file storage (such as NAS and CPFS) in PAI-DLC, you can now configure multiple connections (nconnect). | When you mount Alibaba Cloud file storage, such as NAS and CPFS, for PAI Deep Learning Containers (DLC), you can configure multiple connections (nconnect). This allows for fine-grained control over the number of mount connections, optimizing concurrent access performance from multiple nodes and ensuring the stability of large-scale training tasks. | |
2025-02-07 | EAS supports multi-machine distributed inference. | With the advent of ultra-large-scale Mixture-of-Experts (MoE) models like Qwen-Max and DeepSeek, a single device can no longer handle their vast number of parameters. To address this, EAS introduces a multi-machine distributed inference solution that overcomes hardware limitations and efficiently supports the deployment and operation of ultra-large-scale models. EAS distributed inference supports various parallelism methods, including Pipeline Parallelism, Tensor Parallelism, and Data Parallelism, and is compatible with high-performance inference engine frameworks such as BladeLLM, vLLM, and SGLang. |
January
Release date | Feature | Description | References |
2025-01-21 | DLC supports training timeout alerts. | DLC supports the configuration of training timeout alerts. Customers can customize timeout alert rules for training tasks during the environment preparation, queuing, and running stages. When a rule is triggered, an alert notification is sent, making it easier for customers to monitor for abnormal training progress. | |
2025-01-21 | DLC supports training status notifications. | DLC supports subscriptions to training status notifications. New status events such as queuing, bidding, environment preparation, and running are added. This makes it easier for customers to track training progress and enhances the message notification capabilities of the training service. | |
2025-01-20 | When you submit tasks in DLC, you can now directly mount storage services. | When submitting training tasks in DLC, you can directly select different storage instances in the submission form. Currently, various Alibaba Cloud storage instances are supported, including OSS, General-purpose NAS, Extreme NAS, General-purpose CPFS, and AI Computing CPFS. This lowers the entry barrier and simplifies usage. | |
2025-01-20 | Ray on DLC supports the use of idle resources. | DLC supports submitting Ray tasks using idle resources. This allows customers to run multiple tasks with a single set of resources, enabling resource sharing between tasks and improving resource utilization. | |
2025-01-20 | ArtLab launches industry tool capabilities. | ArtLab has launched industry tool capabilities. The first phase includes applications such as realistic E-commerce product images (home appliances and furniture), corporate-style poster generation, and creative footwear design. More pre-built applications will be continuously added in the future. | |
2025-01-20 | ArtLab launches the AIGC Application Zone. | PAI-ArtLab launches the AIGC Application Zone module. It allows users to use online applications that encapsulate ComfyUI workflows for operations like text-to-image and image-to-image generation. This lowers the barrier to using AIGC production tools and reduces user costs through a serverless service model. 1. Out-of-the-box: Start applications with one click, no environment configuration required. 2. Built-in enterprise-level AIGC applications, such as generating corporate-style posters and creating profile pictures for corporate events. 3. Serverless application mode: Billing occurs only during GPU inference, significantly reducing user costs. | |
2025-01-20 | Model Gallery supports model inference acceleration. | Pre-trained models in PAI-Model Gallery can be matched with supported inference acceleration capabilities (such as vLLM and BladeLLM) based on the selected instance type. | |
2025-01-16 | EAS upgrades to the new BladeLLM high-performance deployment service. | PAI-EAS supports scenario-based deployment of BladeLLM, achieving faster response times and higher throughput for LLM inference. BladeLLM is a self-developed inference engine by PAI that provides an efficient runtime, high-performance operator implementation, and hybrid quantization. PAI-EAS fully integrates with BladeLLM to launch a high-performance LLM inference service. It supports the deployment of pre-built and custom models, and lets you enable advanced options like model parallelism and speculative sampling with a single click, providing an efficient LLM deployment solution. | |
2025-01-02 | Model Gallery is officially launched in multiple regions, including China (Hong Kong). | PAI-Model Gallery integrates pre-trained models from fields such as LLM, CV, NLP, and speech. It provides one-stop, no-code features for model training, model compression, model evaluation, and model deployment. The service is now available in the following new regions: China (Hong Kong), Japan (Tokyo), Indonesia (Jakarta), Germany (Frankfurt), and US (Virginia). |
2024
December
Date | Feature | Description | Reference |
2024-12-23 | DLC pay-as-you-go bills distinguish job types | DLC training jobs support system tags (key:acs:pai:payType) that distinguish pay-as-you-go jobs from preempted jobs. Customers can quickly identify and filter pay-as-you-go jobs and check the consumption and discount. | |
2024-12-16 | Machine Learning Designer supports grouping LLM data preprocessing components | Machine Learning Designer supports grouping multiple serial data preprocessing nodes (DLC) for execution, avoiding repeated data writes to disk and the time consumption of starting and stopping distributed tasks, thereby improving execution efficiency and supporting automatic intelligent aggregation. | |
2024-12-16 | DLC launchs preemptible jobs that use general computing resources | PAI supports preemptible jobs based on general-purpose computing resources to provide customers with more cost-effective AI computing power. | |
2024-12-10 | PAI training service launched in Germany (Frankfurt) | Deep Learning Containers (DLC) and AI resource quotas are available in the Germany (Frankfurt) region. You can submit training jobs that use resource quotas and pay-as-you-go public resources. | |
2024-12-09 | DLC job status upgrades to v2.0 | Based on the resource quota, the Queuing and PreAllocation states are merged into Queuing. This provides clear and simple task status information for easy use and understanding. | |
2024-12-06 | DLC sanity check supports custom items | The DLC sanity check feature supports more than 15 items, such as computing performance checks, node communication checks, computing and communication cross-checks, and model simulation checks. This further improves its troubleshooting capabilities in case of computing power and network failures. The check items are open to users to choose based on their business requirements to achieve self-management and control. |
November
Date | Feature | Description | Reference |
2024-11-20 | DSW supports dynamic mounting of OSS datasets | 1. Allows dynamic mounting or unmounting of OSS datasets without restarting the instance. 2. Provides an easy-to-use SDK that allows users to mount or unmount datasets through simple configuration or a single line of code. 3. Supports dynamically mounting of datasets from AI assets (PAI public datasets or custom datasets) or directly mounting OSS storage paths. | |
2024-11-20 | DSW instances support custom access configurations | With the rapid development of AIGC, various WebUI frameworks and application development frameworks have become the mainstream choice for developers. As a one-stop AI development platform, PAI-DSW provides the custom service access configuration feature. During application development, developers can share services with collaborative developers for testing and verification in a secure manner at any time. |
October
Date | Feature | Description | Reference |
2024-10-17 | AI general computing resource group supports L20 in international regions | AI general computing resource group of PAI supports L20 (gn8is series) in international regions | |
2024-10-12 | DLC job status upgrades to v1.0 | Computing power type includes resource quotas, bidding resources, and public computing power. Business mode includes subscription, bidding, and pay-as-you-go. At job and instance levels, the EnvPreparing and Bidding statuses are added. The Created, Queuing, and PreAllocation statuses are simplified. This provides clearer and simpler job status information for ease of use and understanding. | |
2024-10-11 | ComfyUI serverless available in PAI-ArtLab | ComfyUI serverless is available in the ArtLab toolbox. You can use ComfyUI for text-to-image or image-to-image generation. The serverless mode reduces costs. You are charged only for model inference. | |
2024-10-10 | QuickStart supports DPO and CPT for LLM | PAI Quick Start-Model Gallery provides more complete LLM training capability. On top of the original Supervised Fine-Tuning (SFT), Model Gallery now supports Direct preference optimization (DPO) and Continued Pre-training (CPT). |
September
Date | Feature | Description | Reference |
2024-09-29 | DSW integrates Tongyi Lingma | The AI coding assistant Tongyi Lingma (Personal Edition) is integrated into DSW, offering features such as line or method level code generation, natural language to code, unit test generation, comment generation, code explanation, AI coding chat, and troubleshooting. Users can use the feature without installation and login, experiencing efficient and graceful coding. | |
2024-10-08 | PAI Training Service is available in the China (Hong Kong) and Indonesia (Jakarta) regions | Deep Learning Containers (DLC) and AI resource quotas are available in the China (Hong Kong) and Indonesia (Jakarta) regions. You can submit training jobs that use resource quotas and pay-as-you-go public resources. |
August
Date | Feature | Description | Reference |
2024-11-11 | Judge model feature officially released | The judge model service of PAI uses a fine-tuned LLM based on Qwen2 as a judge to score responses from evaluated models. This service is suitable for open-ended and complex scenarios. Main advantages: 1. Accuracy: The judge model can classify subjective questions into scenarios such as open-ended discussions, creative writing, code generation, and role-playing. It then develop tailored criteria for each scenario, significantly enhancing evaluation accuracy. 2. Efficiency: Without the need for manual data labeling, the judge model can independently analyze and evaluate LLMs based on questions and model answers, greatly boosting evaluation efficiency. 3. Ease of use: PAI offers various usage methods, such as task creation in the console, API calls, and SDK calls. This allows for both quick trials and flexible integration for developers. 4. Cost-effectiveness: The judge model provides performance evaluation at a competitive price. Its performance is comparable to that of ChatGPT-4 in Chinese language scenarios. | |
2024-09-03 | DSW supports NotebookLab (Lightweight Edition) | 1. Lightweight coding is supported in Notebook. You can use a browser to code, without the need to start other resources in advance. 2.Notebooks are managed as assets and are decoupled from instance resources, making it easier to archive or share Notebooks as documents or code samples. | |
2024-08-26 | EAS supports LLM Intelligent Router to improve LLM inference efficiency | When customers deploy LLM services on EAS, they can enable the LLM Intelligent Router feature. LLM Intelligent Router can evenly allocate the computing power and video memory of backend inference instances and improve the resource usage of clusters. | |
2024-08-26 | DLC jobs with general computing resources support CPU affinity | The general computing resources of PAI DLC supports CPU core binding to improve job performance. | |
2024-08-15 | EAS supports dedicated gateway | EAS supports the dedicated gateway feature to meet inference requirements for security isolation and access control. This reduces network risks in high-concurrency and high-throughput business scenarios. Dedicated gateways allow you to configure whitelists for access over virtual private clouds (VPCs) and the Internet and implement fine-grained management. Dedicated gateways can also ensure the stability of connections between services. You can use PrivateLink to connect to the VPC of your enterprise. You can also implement independent control over access to the Internet. | |
2024-08-15 | PAI workspace supports custom roles | Workspace is a key concept in Platform for AI (PAI). Workspaces allow your organization or team to manage computing resources, user permissions, and AI assets in a centralized manner to achieve seamless collaboration at every stage of AI development. In specific scenarios, the current preset roles of a workspace cannot meet the management requirements of customers. For example, you cannot assign a RAM role the permissions to use DSW without assigning the role the permissions on DLC. PAI provides the feature that allows you to customize roles and related permissions. | |
2024-08-05 | Discontinuation of earlier versions of PAI-PyTorch algorithm components | Earlier versions of the PyTorch algorithm components, including PyTorch100 and PyTorch 131-based components, are officially discontinued in all clusters of Platform for AI (PAI) on August 30, 2024 because of a system upgrade. If you have a PyTorch job that is submitted to MaxCompute by using the If you have questions or require technical support, contact us in your dedicated DingTalk group or submit a ticket. Thank you for your cooperation. |
July
Date | Feature | Description | Reference |
2024-07-03 | EAS supports GPU sharing | When you deploy a model in EAS, you can split and use the computing power based on the ratio of GPU computing power and the memory size. This helps reduce resource costs and improve resource utilization. On the deployment page, you can schedule instances based on GPU memory and computing power. This allows multiple instances to share a single GPU. | |
2024-07-03 | EAS supports instance health check | The health check feature of EAS ensures high service availability. By performing fast fault detection and automatic recovery, this feature can facilitate enterprise-level inference service deployment. The Kubernetes health check mechanism can automatically detect and recover failed containers. This ensures that traffic is allocated only to healthy instances. |
June
Date | Feature | Description | Reference |
2024-07-01 | QuickStart supports LLM evaluation | PAI-QuickStart provides the LLM evaluation feature. You can evaluate the comprehensive capabilities of a model based on authoritative public datasets, such as CMMLU, C- Eval, or MMLU, or custom datasets to compare the performance of multiple models and determine whether the model capabilities are suitable for your business scenarios. | |
2024-06-19 | General computing resources of PAI support mounting CPFS for Lingjun (invitational preview) | PAI leverages Alibaba Cloud storage services to provide cost-effective solution for storage and computing in large language model scenarios. You can mount CPFS for Lingjun for a PAI training that runs on general computing resources. | |
2024-06-12 | Machine Learning Designer is available in the China (Ulanqab) region | Machine Learning Designer is available in the China (Ulanqab) region. You can deploy your services in this region in the PAI console. | |
2024-06-11 | Machine Learning Designer provides the Notebook component | Machine Learning Designer provides the Notebook component, which can be connected to DSW instances. The component allows you to directly use notebooks to write, debug, and run code in a pipeline. |
May
Date | Feature | Description | Reference |
2024-07-01 | QuickStart supports fine-tuning LLMs by using QLoRA, LoRA, and full parameter | PAI allows you to fine-tune an LLM by using QLoRA, LoRA, and full-parameter fine-tuning in QuickStart. You can select a training method based on your business requirements to reduce costs. | |
2024-06-07 | DSW supports instance RAM role configuration | After you associate the default role of PAI with an instance, you do not need to configure an AccessKey pair in the following scenarios in DSW:
If you use a custom RAM role, DSW uses the temporary access credentials of the role to access specific Alibaba Cloud services, such as OSS or RDS. This ensures secure communication between the DSW instance and other Alibaba Cloud services. |
April
Date | Feature | Description | Reference |
2024-04-29 | EAS supports serverless deployment of AI painting services | EAS provides the serverless deployment feature for model services that have intermittent or unpredictable traffic patterns. If you deploy an AI painting service by using the serverless method, you are charged only for GPU consumption. | Quickly deploy Stable Diffusion for text-to-image generation in EAS |
March
Date | Feature | Description | Reference |
2024-03-25 | DSW supports integrated AI and big data development | Customers can submit data analysis or preprocessing tasks to MaxCompute or E-MapReduce by using multiple methods such as Python in DSW. The processed data can be used in model trainings that are performed on on-premises GPU devices or in Deep Learning Containers (DLC). | |
2024-03-25 | The file transfer station feature is available in DSW | DSW provides the file transfer station feature that can accelerate the upload process when you need to upload large files such as large models from your on-premises computer to a DSW instance. After you upload the large file, you can use the uploaded large file in multiple DSW instances in your RAM account. | |
2024-03-15 | PAI-Lingjun Intelligent Computing Service available in the Singapore region | PAI-Lingjun Intelligent Computing Service is a next-generation intelligent computing service independently developed by Alibaba Cloud to provide optimized heterogeneous cluster instances. Trained based on a large number of AI applications, PAI-Lingjun Intelligent Computing Service is proven to provide high performance, efficiency, and resource utilization. PAI-Lingjun Intelligent Computing Service can meet the requirements of various industries, such as autonomous driving, basic scientific research, new drug research and development, finance, and meta-universe. The service provides affordable and accessible intelligent computing power to facilitate technological innovation and industrial upgrade. PAI-Lingjun Intelligent Computing Service is available in the Singapore region. You can activate the service in the console. | |
2024-06-06 | Discontinuation of GPU-related server and algorithm components in Machine Learning Designer | The warranty for V100 and P100 server clusters in which services are running is expired. The algorithm components that are related to TensorFlow(GPU), MXNet, and PyTorch in Machine Learning Designer of Platform for AI (PAI) are discontinued on March 1, 2024. You can continue using the related algorithm components of the cloud-native version and submit training jobs to Deep Learning Containers (DLC) in PAI. We recommend that you use the Python-based components in Machine Learning Designer to run DLC jobs. The Python-based components can work in the same manner as the discontinued components. The existing tasks of the discontinued algorithm components are not covered by the SLA starting from June 1, 2024. The V100 and P100 server clusters are discontinued on June 30, 2024. |
February
Date | Feature | Description | Reference |
2024-02-28 | Machine Learning Designer provides data preprocessing operators for LLM and common templates | High-quality data preprocessing is an important step in LLM application. Machine Learning Designer of Platform for AI (PAI) provides commonly used high-performance operators for data preprocessing, such as deduplication, standardization, and sensitive information masking. You can use large-scale distributed computing capabilities based on MaxCompute when you preprocess data for LLMs to improve efficiency and improve the reliability and performance of LLMs. | |
2024-02-04 | EAS serverless deployment is in invitational preview | EAS provides the serverless deployment feature for model services that have intermittent or unpredictable traffic patterns. If you deploy a model service in EAS by using the serverless deployment feature, you are charged only when GPU computing occurs. For example, if you deploy an AI painting model, you are charged based on the actual painting duration. |
January
Date | Feature | Description | Reference |
2024-02-04 | QuickStart Released on the International Site (alibabacloud.com) | The QuickStart feature is now available in the Singapore region. | |
2024-02-04 | EAS simple deployment released | EAS provides simplified deployment methods for common deployment scenarios, including ModelScope model deployment, Hugging Face model deployment, Triton deployment, TFServing deployment, LLM deployment, and SD web application deployment. In these scenarios, you need to only provide the storage directory of the model to start services and applications in a few clicks. | |
2024-02-01 | Quick deployment of AI video generation applications by using EAS | Users can use Elastic Algorithm Service (EAS) to deploy web applications for AI video generation based on ComfyUI and Stable Video Diffusion models. EAS can help you quickly implement AI-powered text-to-video or image-to-video generation in industries such as live streaming and short video platforms, gaming and Internet entertainment, and animation production. |
2023
December
Date | Feature | Description | Reference |
2023-12-13 | Machine Learning Designer is available in Indonesia (Jakarta) | Machine Learning Designer is available in Indonesia (Jakarta). You can select the region Indonesia (Jakarta) on the Platform for AI (PAI) console. | |
2023-12-06 | DSW supports SSH logon | Allows customers to access DSW instances by using the machines in their VPC or from the on-premises development environment in a more convenient manner and facilitates development and training in DSW. |
November
Date | Feature | Description | Reference |
2023-11-20 | PAI releases the automated machine learning platform AutoML | PAI releases AutoML, an enhanced machine learning service of PAI. AutoML integrates various algorithms and distributed computing resources supported by PAI and supports multiple access methods. You can use AutoML to automatically find the optimal hyperparameter values and improve the model tuning efficiency. |
October
Date | Feature | Description | Reference |
2023-10-27 | Subscription AI Training is available on the International site (alibabacloud.com) | AI Training of Platform for AI (PAI) is available on the international site (alibabacloud.com) in the following regions: China (Beijing), China (Shanghai), China (Hangzhou), China (Shenzhen), and Singapore. AI Training supports the subscription billing method. You can access AI Training in the PAI console. |
September
Date | Feature | Description | Reference |
2023-09-28 | EAS supports push-button deployment of the Tongyi Qianwen model | You can use PAI-EAS to deploy web UI applications that are based on the open source Tongyi Qianwen model, and use the web UI and API operations to perform model inference. Qwen-7B is a 7 billion-parameter model of the Tongyi Qianwen series developed by Alibaba Cloud. Qwen-7B is a large language model based on Transformer, and is trained on ultra-large-scale pre-trained data. The pre-trained data covers a wide range of data types, including a large number of texts, professional books, and code. In addition, Qwen-7B-Chat, an LLM AI assistant is developed based on Qwen-7B. | |
2023-09-18 | DLC supports monitoring metric subscription and alert | PAI-DLC offers detailed monitoring metrics for job resource conditions, enabling users to configure flexible alert rules for distributed training jobs. | |
2023-09-18 | EasyCKPT high-performance CKPT released | PAI-EasyCKPT, designed for PyTorch model training, features a near-zero overhead saving mechanism and ensures accurate model saving and recovery throughout the training process. Compatible with Megatron and DeepSpeed, it requires minimal code changes for use. |
August
Date | Feature | Description | Reference |
2023-09-04 | Support for deploying and fine-tuning Stable Diffusion models |
|
July
June
May
Date | Feature | Description | Reference |
2023-05-20 | Support for PAI SDK for Python | PAI SDK for Python provides easy-to-use HighLevel APIs for machine learning engineers to easily train and deploy models on PAI and complete end-to-end machine learning processes. |
April
Date | Feature | Description | Reference |
2023-04-19 | Support for elastic resource pool in Elastic Algorithm Service (EAS) | EAS supports auto scaling of service resources. If node resources in the dedicated resource group are insufficient during a service scale-out, new instances of the service are created in the pay-as-you-go public resource group and billed based on public resource group rules. During a service scale-in, service instances that reside in the public resource group are released first. | |
2023-04-04 | Support for the new version of Elastic Algorithm Service (EAS) page for quick service deployment | EAS supports the following deployment methods: Deploy Service by Using Image, Deploy Web App by Using Image, and Deploy Service by Using Model and Processor. You can deploy AI services or applications to EAS in just a few clicks. |
March
Date | Feature | Description | Reference |
2023-03-23 | Upgraded model management feature | Models trained by Machine Learning Designer can be registered on the model management page. You can change the approval status of model versions to trigger model-related events, such as automated messaging to DingTalk groups through chatbots, or calling the specified HTTP or HTTPS service. |
February
Date | Feature | Description | Reference |
2023-02-13 | Support for preemptible resource instances in Elastic Algorithm Service (EAS) | When you use a public resource group to deploy a service to EAS, you can specify preemptible instances for the service to reduce costs. | |
2023-02-06 | Support for multiple instance types for Elastic Algorithm Service (EAS) | When you deploy services in EAS, you can specify multiple instance specifications in the configuration file. The system then prepares resources based on the instance types you specified in the configuration file. This method reduces the chances of insufficient resources in scenarios where only a single instance type is specified. |
January
Date | Feature | Description | Reference |
2023-01-13 | Support for push-button pipeline deployment as an EAS online service | Machine Learning Designer supports push-button pipeline deployment. You can deploy a batch data-processing pipeline that implements data pre-processing, feature engineering, and model prediction to Elastic Algorithm Service (EAS) as an online service after packaging the pipeline as a model. |
2022
December
Feature | Description | Date | Region | Reference |
Support for computing resources based on Yitian 710 in Elastic Algorithm Service (EAS) | EAS supports computing resources based on Yitian 710 processors. These resources can help you reduce the costs of model deployment, model inference, as well as improve efficiency. | 2022-12-8 | All regions | |
New algorithm components | Machine Learning Designer provides a variety of new algorithm components, including Prophet, MTable Expander, MTable Assembler, and Time Window SQL. You can find and use the components in the left-side directory tree of components on the Machine Learning Designer platform. | 2022-12-5 | All regions | |
Support for custom templates | You can create a custom template based on a pipeline that runs successfully in Machine Learning Designer. You can use this template to quickly build similar pipelines to improve efficiency. | 2022-12-1 | All regions |
December
Date | Feature | Description | Applicable customers | Reference |
Support for O&M on EAS nodes | You can perform O&M operations on the nodes in resource groups. The operations include viewing node information, stopping and restarting node scheduling, and removing instances in nodes. | 2022-11-30 | All regions | |
Support for querying the updates of Data Science Workshop (DSW) instances | You can view the changes to the states of a DSW instance throughout the lifecycle of the instance. You can view the details of a DSW instance and change its configurations. | 2022-11-18 | All regions |
September
Feature | Description | Date | Region | Reference |
Support for the service grouping and asynchronous inference features in Elastic Algorithm Service (EAS) | When you create an EAS service, you can specify the service group to which the EAS service belongs. The service group has a unified ingress. The ingress allocates traffic to each EAS service based on the traffic allocation policy. You can also specify the traffic allocation ratio of each service in a service group to ensure high resource utilization. PAI provides the queue service and asynchronous inference features. These features allow you to consume inference services by distributing requests, subscribing to requests and pushing inference results, or periodically querying inference results. | 2022-09-30 | All regions |
November
Date | Feature | Description | Applicable customers | Reference |
New algorithm components | Machine Learning Designer provides a variety of new training and prediction components, including XGBoost, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Gaussian Mixture Mode (GMM), Ridge Regression, and Lasso Regression. You can find and use the components in the left-side directory tree of components on the Machine Learning Designer platform. | 2022-08-02 | All regions |
July
Feature | Description | Date | Region | Reference |
Python Script V2 added | The Python Script V2 component is added to Machine Learning Designer of PAI. You can use the component to develop custom algorithms. You can also use the component together with the pre-set algorithms of PAI to support more scenarios. | 2022-07-15 |
| |
Designer officially launched in US (Virginia) | Designer is officially launched in US (Virginia). You can select the corresponding region in the PAI console and create a workspace to use Designer-related features. | 2022-07-05 | US (Virginia) | None |
Automatic stress testing available | The automatic stress testing feature is available for EAS. You can use EAS-benchmark, a distributed stress testing tool, to create stress testing tasks for prediction services that are deployed in EAS. | 2022-07-04 |
| N/A |
October
Date | Feature | Description | Applicable customers | Reference |
Visualized analytical reports supported | Analytical reports can be visualized by using Tensorboard. Visual deep learning components of Machine Learning Designer allow you to use the Tensorboard dashboard to view visualized analytical reports. On the dashboard, you can view visualized feature importance evaluation, correlation analysis, and scatter charts. | 2022-06-22 |
| |
Designer officially launched in China (Hong Kong) | Designer officially launched in China (Hong Kong), offering hundreds of PAI self-developed machine learning algorithms and dozens of industry templates. You can use them on the PAI console as needed. | 2022-06-20 | China (Hong Kong) | None |
May
Feature | Description | Publish date | Publish region | Reference |
Available in Singapore and US (Silicon Valley) | Machine Learning Designer is available in the Singapore and US (Silicon Valley) regions. Machine Learning Designer provides hundreds of self-developed machine learning algorithms and dozens of industry templates. You can use them as needed in the PAI console. | 2022-05-10 |
| None |
April
Feature | Description | Date | Region | Reference |
Fully-managed Flink resources supported | Fully-managed Flink resources can be purchased and associated with workspaces. Then, you can use multiple components or use the PyAlink Script component alone to build pipelines for large-scale distributed training of models. | 2022-04-30 | Germany (Frankfurt) | |
New anomaly detection, recommendation, data source, and custom algorithm components added | Components including PyAlink Script, Read CSV File, IForest Outlier, LOF Outlier, One-Class SVM Outlier, and Swing Recommendation are added to Machine Learning Designer. The PyAlink Script component allows you to call hundreds of algorithms that are under the Alink framework. | 2022-04-16 | Germany (Frankfurt) |
September
Date | Feature | Description | Applicable customers | Reference |
Designer officially launched in Germany (Frankfurt) | Designer officially launched in Germany (Frankfurt), offering hundreds of PAI self-developed machine learning algorithms and dozens of industry templates. You can use them on the PAI console as needed. | 2022-03-30 | Germany (Frankfurt) | None |
TensorFlow 2.7 supported by PAI-Blade | TensorFlow 2.7 is supported by PAI-Blade. You can select a version based on your needs. | 2022-03-27 | All regions | None |
DSW is officially launched in five regions including Singapore | You can create DSW instances and use DSW features to build and train models in these regions. | 2022-03-21 |
| None |
CronHPA feature supported, and gRPC and WebSocket protocols supported for image deployment and release | The CronHPA feature is available in EAS. This feature allows you to perform scheduled auto-scaling on service instances. In addition, EAS supports the deployment of services by using the open source TensorFlow Serving system or the Triton software. | 2022-03-21 |
|