How QwQ-32B Outperforms OpenAI o1-mini and Deploys in One Click on Alibaba Cloud

01 QwQ-32B Model Introduction

On March 6, Alibaba Cloud released and open‑sourced the inference model Tongyi Qianwen QwQ‑32B. Through large‑scale reinforcement learning, QwQ‑32B achieves a qualitative leap in mathematics, code, and general abilities, with overall performance comparable to DeepSeek‑R1 while reducing deployment cost.

In a series of authoritative benchmark tests, QwQ‑32B outperforms OpenAI o1‑mini and matches the strongest open‑source inference model DeepSeek‑R1. On the AIME24 math set and LiveCodeBench code set, its scores are on par with DeepSeek‑R1 and far exceed o1‑mini and distilled R1 models of the same size. In Meta‑led “Hardest LLMs” LiveBench, Google’s IFEval, and UC Berkeley’s BFCL function‑calling tests, QwQ‑32B scores surpass DeepSeek‑R1. The model also integrates agent‑related capabilities for tool use, critical thinking, and adaptive reasoning.

02 PAI‑Model Gallery Introduction

PAI‑Model Gallery, a component of Alibaba Cloud’s AI platform PAI, aggregates high‑quality pretrained models from global open‑source communities across LLM, AIGC, CV, NLP, and other fields. It enables zero‑code end‑to‑end training, deployment, and inference, simplifying AI development for developers and enterprises.

Gallery URL: https://x.sm.cn/Hj94lkS

03 One‑Click Deployment of QwQ‑32B

Enter the Model Gallery page (link above).

Log in to the PAI console, select the appropriate region (all regions except Beijing support QwQ‑32B).

Choose the workspace, then go to “Quick Start > Model Gallery”.

In the model list, click the QwQ‑32B model card to open its detail page.

Click “Deploy”, select a deployment framework (SGLang, vLLM, BladeLLM), configure the inference service name and resources, and deploy to the PAI‑EAS inference service platform.

After deployment, view the endpoint and token on the service page; refer to the model’s documentation for invocation details.

You can also debug the deployed QwQ‑32B service online on PAI‑EAS; the model demonstrates strong chain‑of‑thought capabilities.

04 Contact Us

Stay tuned to PAI‑Model Gallery for new SOTA models. For model requests, contact us via DingTalk (group 79680024618) or scan the QR code below.