Alibaba’s response to Deepseek; The QWQ-32B model was introduced with 32 billion parameter

Chinese company Alibaba Of his latest open source linguistic model named QWQ-32B Unveiled; A 32 billion parameter model aimed at improving the ability to solve complex problems and logical reasoning. This model uses reinforcing learning (RL) and advanced techniques in areas such as Mathematics, Coding And Complicated issues analysis Offers.

According to reports, the QWQ-32B is an advanced version of the QWQ, which Alibaba had released in November 2024 with the aim of competing with the O1-Preview Model of Openai. At the outset, this model attracted much attention due to the optimal performance in mathematical tests (Aime, Math) and Scientific reasoning (GPQA); However, in the field of programming, competitors such as LiveCodebench are lagging behind.

The QWQ-32B is now trying to address these weaknesses, relying on the multi-stage learning structure. According to preliminary results, this model has been able to get closer to the performance level of large models like DeEPSEK-R1 with 671 billion parameters, while only 24GB of GPU memory; Deepseek-R1, however, needs more than 1500 GB of VRAM.

Technical and architectural specifications of the QWQ-32B artificial intelligence model

The QWQ-32B model has the following features:

• 64 layers of transformer with techniques such as Rope and Swiglu
• Support of 131,072 token for long text processing
• Generalized Architecture of Query Attention (GQA)
• Three -step tutorials including pre -development, monitoring and reinforcement learning

Also, the QWQ-32B reinforcement learning is implemented in two steps; First, focusing on precision in mathematics and programming and then improving public abilities such as understanding instructions and coordinating human behavior.

This model can be a good option for companies seeking to implement automatic data analysis, software development, financial modeling, or customer service automation because of its open source and advanced reasoning. Also, although some non -perishable users may have concerns about the security and bias of Alibaba -affiliated models, the release of this model in Hugging Face for download and use offline greatly reduces these concerns.

The QWQ-32B model is released under the Apache 2.0 license and is available through Hugging Face and Modelscope platforms. This allows companies and developers to use it to produce products, services and even monetary projects without the restrictions of commercial models.

It is also applicable to the QWen Chat service. The QWen team plans to pave the way for artificial intelligence (AGI) by continuing to develop this model.