All Products
Search
Document Center

Function Compute:Release notes in 2025

Last Updated:Jan 09, 2026

This topic describes the feature releases and documentation updates for Function Compute (FC).

December 2025

Feature name

Change type

Description

References

GPU hybrid instance mode

(Available in Shenzhen and Shanghai)

New

Supports a hybrid mode with resident and on-demand instances. Resident instances handle stable traffic. On-demand instances automatically scale up for peak loads. Idle on-demand instances enter a low-cost state. This reduces costs while maintaining stability and availability.

All GPU types now support request-based auto scaling

New

All GPU types now support auto scaling based on request volume. The system dynamically creates and releases instances based on real-time concurrent requests. This matches resource consumption to your business load and offers a flexible cost-saving solution for non-persistent tasks. (Note: New instances experience cold starts. For latency-sensitive services, configure provisioned instances.)

GPU image size limit increased to 30 GB

Optimization

The maximum size for compressed images used in GPU function deployments is increased from 15 GB to 30 GB. This improves the deployment success rate for images with large dependencies or Large Language Models (LLMs). It also reduces the cost of splitting images and using multi-stage builds.

Quotas and Limits

PolarDB dynamic mounting

(Available in Hong Kong (China), US West, and Singapore)

New

Supports dynamic mounting of PolarDB storage directories. This lets different instances share persistent data and large model weight files. It improves data access performance and optimizes the delivery and O&M efficiency for complex enterprise applications.

OSS dynamic mounting

(Available in Hong Kong (China), US West, and Singapore)

New

Dynamically mount or unmount OSS Bucket directories at runtime without restarting services or rebuilding images. This feature decouples computing from large-scale unstructured data. It simplifies loading model training data and streaming multimedia content.

Native CORS policy configuration for HTTP triggers

New

Configure native Cross-Origin Resource Sharing (CORS) rules. Customize allowed origins, request methods, headers, and preflight request cache times as needed. This lets you securely handle cross-domain calls from browsers without writing extra middleware code.

Console Overview page upgrade and FunctionQ assistant

Optimization

The Function Compute (FC) Overview page is redesigned to include Cost Center, Resource Management, and Solutions+AI sections. The new FunctionQ intelligent development assistant provides developers with precise document indexing and technical support.

November 2025

Feature name

Change type

Description

References

GPU integration with ARMS to support LLM observability

New

Deploy an OpenTelemetry Collector in sidecar mode within a function instance. This automatically collects and reports Prometheus metrics exposed by the LLM inference engine, such as token throughput and request latency. This enables production-grade, visual, and alert-ready real-time monitoring for LLM services.

Integration solution for monitoring LLM inference model service metrics

DevPods are supported on custom image runtimes.

New

DevPod provides a complete cloud-based container development environment. It integrates VS Code, Jupyter, and a terminal. DevPod supports custom images, persistent storage, and remote debugging. It is ideal for AI model development.

DevPod

Elastic policy for resident resource pools

New

GPU resident resource pools now support request-based elasticity, scheduled elasticity, and metric-based scaling to improve resource utilization.

Instance scaling limits and elastic policies

Ulanqab compliance cloud launch

New

Launched dedicated cloud and general industry cloud in the Ulanqab region.

Available regions

Service endpoints

Delayed release for on-demand instances supports light hibernation

New

Both CPU and GPU on-demand instances support the delayed release feature. In delayed release mode, both instance types support billing for light hibernation. The delayed release feature for GPU on-demand instances does not currently support light hibernation for WebSocket connections.

Configure delayed release for on-demand instances

October 2025

Feature name

Change type

Description

References

Delayed release for on-demand instances

New

For background asynchronous tasks, such as uploading logs or synchronizing data after a request is processed, use the delayed release feature. This prevents task interruptions and data loss. After you set a delayed release time, the system automatically keeps the on-demand instance active, switches it to an idle state, or destroys it if no requests are received after a timeout. This is based on vCPU utilization.

Configure delayed release for on-demand instances

Dynamic NAS mounting

New

Use the dynamic NAS mounting feature to dynamically mount each user session of a function instance to a dedicated subdirectory in File Storage NAS. This provides secure isolation for persistent data in a multi-tenant environment.

Dynamically mount File Storage NAS

Support for new GPU types

New

Added support for two new GPU types: Hopper.2 and Blackwell.1.

Instance types and specifications

September 2025

Feature name

Change type

Description

References

gRPC support for HTTP triggers

New

Function Compute supports the gRPC protocol. You can directly trigger gRPC services using HTTP triggers. The function acts as a gRPC server to process streaming and non-streaming client requests. It also benefits from the elastic scaling and fully managed capabilities of the serverless architecture.

MCP Streamable affinity

New

Enable the MCP Streamable HTTP affinity feature to route requests from the same MCP session to the backend function instance that created the session. When an MCP Streamable HTTP session is initialized, Function Compute parses the Mcp-Session-Id field in the HTTP header of the function's response. It then associates this ID with the session on the platform. Subsequent requests with the same Mcp-Session-Id are considered part of the same MCP Streamable HTTP session and are routed to the instance that initialized the session.

Configure MCP Streamable HTTP affinity

Session lifecycle management

New

To improve session manageability, Function Compute introduces session API management for HeaderField and Cookie affinity. This upgrades sessions to explicit resources and provides a complete set of lifecycle management APIs. You can create, query, update, and delete sessions.

Session lifecycle management

August 2025

Feature name

Change type

Description

References

Tracing Analysis feature

New

Function Compute now integrates with Alibaba Cloud Managed Service for OpenTelemetry. This service is based on the W3C protocol of the de facto standard OpenTelemetry. It helps you easily identify and diagnose performance bottlenecks in distributed applications. This significantly improves development and O&M efficiency in a serverless architecture.

Configure Tracing Analysis

Basic authentication

New

Enable Basic authentication for a custom domain name or an HTTP trigger. The client must include Base64-encoded credentials of the configured username and password in the request. Access to the bound function service is granted only after the credentials are verified.

July 2025

Feature name

Change type

Description

References

Session affinity

New

The session affinity feature ensures that requests from the same client are routed to the same function instance for a period of time. Its core purpose is to maintain session continuity, especially in stateful services.

Instance isolation

New

To meet the demand for isolated runtime environments for AI Sandboxes, Function Compute introduces instance isolation. With this feature, an instance can only process a single request or all requests within a single session. The instance is released when the request or session ends. This provides a secure, reliable, elastic, and cost-effective sandbox runtime environment.

Integrating FC with resource groups

New

To manage Function Compute resources more efficiently, use resource groups to group your resources. Resource groups let you group resources by dimensions such as department, project, or environment. In combination with Resource Access Management (RAM), you can achieve resource isolation and fine-grained permission management within a single Alibaba Cloud account.

Configure resource groups

Resident resource pools (subscription)

New

Purchase resident resource pools to lock in computing power resources of a specific specification in advance. You can then assign a specific number of resident instances of a certain type to your functions as needed. This ensures business stability and provides fixed, controllable costs.

May 2025

Feature name

Change type

Description

References

New command-type Initialize hook

New

The Initialize hook now supports a command type in addition to the code invocation type. You can configure a command to send an HTTP request to the function for initialization after the instance starts.

A typical scenario is prefetching a model in an AI inference application by configuring a prefetch script in the Initializer hook.

Currently, only GPU functions support the command-type Initializer hook.

Bearer authentication

New

In Bearer authentication, you configure the token information that is allowed to access the function in the Function Compute console. When a client initiates a request, it carries a valid token in the Authorization header. The function can be accessed only if the token in the request matches the token configured on the trigger.

April 2025

Feature name

Change type

Description

References

Enhanced function management

Optimization

The function list in the Function Compute console is optimized in the following ways to help you efficiently manage large-scale function resources.

  • Function display: The function list now includes columns for function tags and specifications. You can sort functions by these new columns. You can also customize columns and use flexible paging to display functions.

  • Function search: You can search for functions by name, description, runtime, specification, and tags. You can also combine search criteria.

  • Batch operations: You can edit function tags in batches.

Configure tags

Function Compute 3.0 support for OSS object access points

New

With an access point, an OSS GetObject event can automatically trigger a function to execute. The result is then returned to the application, enabling automated data processing and business workflows.

Trigger a function with a GetObject request using an Object FC Access Point

Function Compute 3.0 support for ZIP file decompression

New

Function Compute supports ZIP file decompression triggers. After you configure decompression rules, uploading a ZIP file that matches the rules to an OSS Bucket automatically triggers a function. The function decompresses the file and writes the extracted files to the target directory.

Configure a ZIP file decompression trigger

March 2025

Feature name

Change type

Description

References

ALB trigger

New

Function Compute supports using an Application Load Balancer (ALB) as an event source. Add a server group of the Function Compute type to an ALB instance to forward requests from the ALB to Function Compute and invoke a function.

ALB trigger

February 2025

Feature name

Change type

Description

References

gRPC protocol support

New

After you bind a domain name to a function, you can use a gRPC client to send gRPC requests. The function acts as a gRPC server that can process gRPC requests and return the results to the client.

gRPC protocol support for custom domain names