Discover QwQ-32B: A 32B LLM Matching 671B DeepSeek‑R1 Performance

On March 6, Alibaba Cloud announced the QwQ-32B inference model, which has only 32 B parameters but achieves performance comparable to the 671 B DeepSeek‑R1 model. In addition to its compact size, QwQ-32B incorporates agent‑related abilities, enabling critical thinking and adaptive reasoning when using tools.

The model is available on Hugging Face, ModelScope, and Ollama, with the following links:

https://huggingface.co/Qwen/QwQ-32B

https://modelscope.cn/models/Qwen/QwQ-32B

https://ollama.com/library/qwq

QwQ-32B is released under the Apache 2.0 license, allowing developers to use it freely and also try it directly via Qwen Chat.

Benchmark results show that QwQ-32B matches DeepSeek‑R1 on the AIME24 math reasoning set and LiveCodeBench programming tests, surpasses o1‑mini and distilled R1 models of similar size, and even exceeds DeepSeek‑R1 on the LiveBench, IFEval, and BFCL evaluations.

Quick Local Deployment

To deploy the model locally with Ollama, follow these two steps:

Install Ollama: curl -fsSL https://ollama.com/install.sh | sh Run QwQ-32B:

ollama run qwq

For macOS or Windows, you can also download the Ollama client from https://ollama.com/download.

Spring AI API Integration

Since Ollama serves QwQ-32B, Java developers can integrate the model into their applications using Spring AI Ollama. The integration process is the same as described for DeepSeek‑R1, simply replacing the model name with qwq.