Deploy the QwQ-32B inference model with one click-Function Compute(FC)-阿里云帮助中心

The QwQ-32B reasoning model has been officially released and open-sourced, quickly gaining significant global attention for its excellent performance and wide range of applications. Powered by Alibaba Cloud Function Compute (FC), Function AI now provides two deployment methods to help you deploy QwQ-32B series models: model service and application template. You can deploy the application template with one click to chat with the model or call the model via an API to integrate it into your AI applications. We invite you to experience QwQ-32B.

QwQ-32B: Smaller size with performance comparable to top open-source models

QwQ-32B was evaluated on a series of benchmarks, including mathematical reasoning, programming, and general capabilities. The following results compare the performance of QwQ-32B with other leading models, including DeepSeek-R1-Distilled-Qwen-32B, DeepSeek-R1-Distilled-Llama-70B, OpenAI-o1-mini, and the original DeepSeek-R1-671B.

On the AIME24 evaluation set for mathematical ability and the LiveCodeBench for coding ability, Qwen QwQ-32B performed on par with DeepSeek-R1-671B. It significantly outperformed OpenAI-o1-mini and R1 distilled models of the same size. Qwen QwQ-32B also scored higher than DeepSeek-R1-671B on several other benchmarks. These include LiveBench, the "hardest LLM evaluation leaderboard" led by Meta's chief scientist Yann LeCun. Other benchmarks include IFEval, an evaluation set for instruction-following ability from Google and others, and BFCL, a test for accurately calling functions or tools from UC Berkeley and others.

Prerequisites

The resources consumed by the GPU function that you create in Function Compute for this tutorial are billed based on the function's specifications and execution duration. If there are no requests, you are charged only for the provisioned snapshot in the shallow hibernation (formerly idle) reserved mode. The Function AI Extreme Speed mode uses provisioned instance snapshots to achieve millisecond-level response times. This mode is technically equivalent to the shallow hibernation (formerly idle) reserved mode of Function Compute, which is suitable for scenarios that require fast cold starts. We recommend that you claim the trial quota for Function Compute to cover resource consumption. Usage that exceeds the trial quota is automatically billed on a pay-as-you-go basis. For more information, see Billing overview.

Method 1: Application template deployment

1. Create a project

Log on to the Function Compute console. In the navigation pane on the left, click Function AI. On the Function AI page, select Projects from the navigation bar, and then click Create Project. Select Create from Template.

2. Deploy the template

In the search bar, enter QWQ. In the search results, click Build an AI chat assistant based on the Qwen-QwQ inference model to go to the Template Details page. Click Deploy Now.
Select a Region and click Deploy Project. In the Project Resource Preview dialog box, you can see the related billable items. For more information, see Products involved in billing. Click Confirm Deployment. The deployment takes about 10 minutes. A status of Deployed indicates that the deployment is successful.
Note
- When you select a region, select the nearest one. If you have enabled a NAS file system and choose to manually configure model storage, select the same region as your file system.
- If a deployment error or model pull failure occurs during testing, this may be because of insufficient GPU resources in the current region. You can try deploying again in a different region.

3. Verify the application

After the deployment is complete, click the Open-WebUI service. Under Endpoint, find Public Network Access and click it. In the Open-WebUI interface, you can chat with the QwQ model.

Method 2: Model service deployment

Use API calls to invoke the model and integrate it into online business applications.

1. Create a blank project

Log on to the Function Compute console. In the navigation pane on the left, click Function AI. On the Function AI page, select Projects from the navigation bar, and then click Create Project.
Select Create Blank Project. In the dialog box that appears, enter a Project Name and Project Description, and then click Create.
On the project details page, click New Service in the upper-left corner. Select Model Service to go to the service configuration page.

2. Deploy the model service

Select the QwQ-32B-GGUF model. Currently, this model is only supported in the China (Hangzhou) region.
Click Resource Configuration. For QwQ-32B-GGUF, the Ada series is recommended. You can use the default configurations. You can also specify the card type and specifications as needed.
Click Preview and Deploy. In the Service Resource Preview dialog box, you can see the related billable items. For more information, see Products involved in billing. Click Confirm Deployment. This stage involves downloading the model and takes about 10 to 30 minutes.

3. Verify the model service

Click Debug to test and verify model calls.

Verify the model call in a local command line window.

4. Third-party platform API calls

You can also verify and use model calls on third-party platforms such as Chatbox. The following example uses Chatbox.

Delete the project

Follow these steps to delete the application and reduce costs.

Go to the Project Details page and click Delete in the upper-right corner. The delete confirmation dialog box appears.
You can see the resources that will be deleted. By default, Function AI selects all services in the project for deletion. To keep a specific service, clear its checkbox. Only the services that remain selected are deleted along with the project.
Select I understand that deleting this project and the selected services will immediately interrupt the online services they support. This action is irreversible and will permanently delete all dependent cloud product resources. Then, click Confirm Deletion.