Deploy Alibaba Cloud’s QwQ-32B LLM: Benchmarks, Agent Features, and One‑Click Setup

QwQ-32B Model Overview

On March 6, Alibaba Cloud released and open‑sourced the new inference model Tongyi Qianwen QwQ-32B. Leveraging large‑scale reinforcement learning, the model achieves a qualitative leap in mathematics, coding, and general abilities, matching the performance of DeepSeek‑R1 while significantly reducing deployment costs.

Benchmark results show that QwQ-32B almost completely surpasses OpenAI o1‑mini and rivals the strongest open‑source inference model DeepSeek‑R1. In the AIME24 math evaluation set and the LiveCodeBench coding benchmark, QwQ-32B performs on par with DeepSeek‑R1 and far ahead of o1‑mini and comparable distilled models. In the “hardest LLMs evaluation” LiveBench led by Meta’s chief scientist, the IFEval instruction‑following benchmark from Google, and the BFCL function‑calling test from UC Berkeley, QwQ-32B scores exceed DeepSeek‑R1. Additionally, the model integrates agent capabilities, enabling critical thinking while using tools and allowing inference to adapt based on environmental feedback.

PAI‑Model Gallery Introduction

The Model Gallery is a component of Alibaba Cloud’s AI platform PAI. It aggregates high‑quality pre‑trained models from domestic and international open‑source communities across LLM, AIGC, CV, NLP, and other domains. By adapting these models through PAI, users can achieve the entire workflow—from training to deployment to inference—without writing code, streamlining AI development for developers and enterprises.

The platform offers flexibility and strong technical support, supporting multiple advanced deployment frameworks:

SGLang provides a simplified configuration for rapid model deployment.

vLLM optimizes large‑scale language models for faster inference.

BladeLLM, Alibaba Cloud’s self‑developed high‑performance inference framework, delivers efficient deployment and inference for massive models.

One‑Click Deployment of QwQ-32B via PAI‑Model Gallery

1. Access the Model Gallery page (https://pai.console.aliyun.com/?regionId=cn-hangzhou#/quick-start/models), log in, select the appropriate region (all regions except Beijing support QwQ‑32B), and choose a workspace.

2. In the Model Gallery’s model list, locate the QwQ‑32B model card and open its detail page.

3. Click the Deploy button in the top‑right corner, select a deployment framework (SGLang, vLLM, or BladeLLM), configure the inference service name and resource specifications, and confirm to deploy the service to the PAI‑EAS inference platform.

4. After deployment succeeds, the service page provides the endpoint and token. Click “View Call Information” to retrieve them, and refer to the model’s documentation for usage instructions.

The deployed QwQ‑32B service can also be debugged online on the PAI‑EAS platform, where its responses demonstrate strong chain‑of‑thought capabilities.