Ali QwQ-32B is released, it may be the best alternative to 671B

Written by
Audrey Miles
Updated on:July-13th-2025
Recommendation

Alibaba Cloud's latest inference model QwQ-32B is coming, with performance comparable to 671B, extremely low resource usage, and open source and free for commercial use!

Core content:
1. QwQ-32B model parameters, performance comparison with DeepSeek-R1
2. QwQ-32B open source protocol, download method and deployment advantages
3. The impact of QwQ-32B release on the industry and the development trend of future AI big models

Yang Fangxian
Founder of 53AI/Most Valuable Expert of Tencent Cloud (TVP)
March 6, 2025 is a memorable day. Alibaba Cloud Tongyi Qianwen official microblog announced the release and open source of the latest inference model QwQ-32B. It is reported that this is a model with 32 billion parameters, and its performance is comparable to DeepSeek-R1 with 671 billion parameters (37 billion of which are activated). Let's take a look at the official content screenshots:
Currently, QwQ-32B has been open sourced on Hugging Face and ModelScope, and has adopted the Apache 2.0 open source protocol, which means that everyone can download it for free and use it for commercial purposes.
Hugging Face also updated the ranking of large models, and QwQ-32B also successfully topped the list:
At the same time, QwQ-32B's overall performance in mathematics, code and general capabilities is comparable to DeepSeek-R1, and it reduces the cost of deployment and use, and can also be deployed locally on consumer-grade graphics cards.
Friends who want to try it can download it through the link below:

huggingface:

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

hf mirror site:

http://hf-mirror.com/Qwen/QwQ-32B

ollama:

https://ollama.com/library/qwq

You can also use ollama to directly pull the model, the command is as follows:

ollama run qwq

Compared with the full version of deepseek-r1 671b, QwQ occupies very little resources! The q4 quantitative version of the ollama platform only occupies 20G of video memory, and the video memory required for the full model does not exceed 80G to achieve smooth reasoning! In addition, there is a rumor that DS related personnel provide technical support!

After the release of QwQ-32B, the impact is still very large, especially for customers who spent a lot of money on 671B before, it may hurt for a while! Now the update iteration of AI big models is too fast. Since DeepSeek was open sourced, it has indeed promoted the leap-forward iteration update of the entire big model. I look forward to these leading companies bringing us more and better technologies!