The new AI model, QWQ-32B-PREVIEW appeared as a strong claimant in the field of AI reasoning, especially since it is available under Apache 2.0, i.e. open for commercial use. Developed by the QWEN team in Alibaba, this model of parameters 32.5 billion can process hints up to 32,000 words and exceeded O1-Review OpenAI and O1-Mini in some references.
According to the testing of Alibaba, the QWQ-32B P-Review exceeds the O1-Review OpenAI model in Aime and Math tests. Aime evaluates models using other AI systems, while mathematics consists of a set of difficult verbal problems. The possibilities of reasoning the new model allow him to solve logical puzzles and solve moderately difficult mathematical problems, although this is not without restrictions. For example, Alibaba admitted that the model can unexpectedly change languages, imprison repetitive loops or struggle with tasks that require a strong justification for common sense.
Unlike many traditional AI systems, QWQ-32B Preview contains the form of a self-esteem mechanism that helps to avoid common mistakes. Although this approach increases accuracy, it also increases the time required to create solutions. Like the OPENAI O1 models, QWQ-32B-PREVIEW uses a systematic reasoning process, planning his steps and performing them methodically to obtain an answer.
Qwq-32b preview to Available on the Hulging Face platformwhere you can download and use it. The model's approach to sensitive topics is in line with other reasoning models, such as the recently released Deepseek, which affect the Chinese regulatory framework. Since companies such as Alibaba and Deepseek operate in accordance with the strict online regulations of China, their AI systems are aimed at compliance with the guidelines promoting “basic socialist values”. This affects the way of responding models to politically sensitive queries. For example, when asked about the status of Taiwan, QWQ-32B-PREVIEW He provided an answer in accordance with the position of the Chinese government. Similarly, the hints about Tiananmen Square resulted in a reaction, reflecting the regulatory environment in which these systems are developed.
While QWQ-32B-PREVEW is sold as available on the basis of a permissible license, not all model elements have been released. This partial openness limits the ability to fully repeat the model or obtain a comprehensive understanding of its architecture. There is a debate on what is “openness” in AI, with models from completely closed systems, offering only API access, to fully open systems revealing all details, including weight and data. QWQ-32B Preview occupies the middle ground in this spectrum.
The increase in reasoning models, such as QWQ-32B-PREVIEW, appears at a time when traditional “provisions on scaling” of artificial intelligence are questioned. Over the years, these provisions suggest that increasing data and calculation of resources would lead to a continuous improvement of AI's capabilities. However, recent reports indicate that the progress rate of models from the leading AI laboratories, including Opeli, Google and Anthropic, began to the plateau. This caused the search for innovative approaches to the development of AI, including new architecture and technology.
One of these approaches to adhesion is the calculation of the test time, also known as the calculation of the application. This method allows AI models to use additional processing time during tasks, increasing their ability to cope with complex challenges. Calculation of test time is the foundation of models such as O1 and QWQ-32B Preview, reflecting the change in focusing on performance optimization during application, and not just relying on training.
The main AI laboratories, apart from Opeli and the Chinese companies, also intensively invest in reasoning models and calculation of test time. The recent report emphasized that Google has significantly expanded his team dedicated to models of reasoning, increasing it to about 200 members. In addition, the company has allocated significant calculation resources to develop this AI research area, signaling the growing involvement of the industry in the future of AI reasoning.