Deploy Open-Source LLMs on Alibaba Cloud Function Compute in 10 Minutes

Solution Overview

The solution leverages Alibaba Cloud Function Compute to deploy open‑source large models from the ModelScope community, enabling a text‑generation service without the need to own or maintain expensive GPU resources. Deployment can be completed in about ten minutes.

Technical Architecture

Function Compute: provides the serverless runtime for serving the LLM.

NAS file storage: stores the ModelScope model files.

VPC (private network): allows Function Compute to access the NAS securely.

Deploy the Application

Open the Function Compute application template, select the ModelScope model source, set Model ID to ZhipuAI/chatglm3-6b, version v1.0.2, region East China 2 (Shanghai), task type chat, and provide your ModelScope Access Token. Keep other options at default and click “Create Application”. Model download may take around 15 minutes.

After the app is created, enable idle‑reserved‑instance mode to avoid cold‑start latency. In the function details page, go to the Configuration tab, choose Reserved Instances, and create a policy with version LATEST, reserved instance count 1, and idle mode enabled.

(Optional) If no NAS is pre‑configured, the system will automatically create or attach a NAS named Alibaba-Fc-V3-Component-Generated. You can bind an existing NAS via the network and storage settings for more control.

Using the LLM Service

When deployment finishes, navigate to the Environment Details page and click the provided domain (e.g., ***.devsapp.net) to access the service. Submit a text prompt and click “Submit”. The first invocation may take ~90 seconds due to cold start; subsequent calls are much faster. The platform automatically returns idle instances to standby after each request.

Important Notes

Do not expose the demo domain publicly to avoid unexpected charges. The domain is intended for learning and testing only; for production use, bind a custom domain as described in the documentation.

Conclusion

This guide demonstrates how to launch an open‑source LLM on Alibaba Cloud Function Compute with minimal setup, leveraging ModelScope models, serverless infrastructure, and reserved‑instance configuration to achieve low‑latency, cost‑effective AI services.