A new AI model, QWQ-32B-PREVIEW, has become a solid competitor in the field of AI reasoning, especially since it is available under a APACHE 2.0 license, that is to say open for commercial use. Developed by the Qwen team from Alibaba, this 32.5 billion parameter model can treat prompts up to 32,000 words and has surpassing O1-PREVIEW from Openai and O1-Mini on certain bearings.
According to Alibaba's tests, QWQ-32B-PREVIEW surpasses the O1-PREVIEW model of OPENAI on love and mathematics tests. Likes to assess models using other AI systems, while mathematics consist of a collection of difficult words problems. The reasoning capacities of the new model allow it to tackle logical puzzles and solve moderately difficult mathematical problems, although it is not without limits. For example, Alibaba has recognized that the model can change my tongue unexpectedly, be trapped in repetitive loops or fight with tasks requiring a strong common sense reasoning.
Unlike many traditional AI systems, the QWQ-32B-PREVIEW includes a form of self-truth mechanism which helps it avoid current errors. Although this approach improves precision, it also increases the time necessary to produce solutions. Similar to the O1 models of OPENAI, QWQ-32B-PREVIEW uses a process of systematic reasoning, planning its steps and executing them methodically to derive responses.
QWQ-32B-PREVIEW is Accessible on the embraced facial platformwhere it can be downloaded and used. The approach of the model on sensitive subjects is aligned with other models of reasoning such as the recently published Deepseek, which are both influenced by Chinese regulatory frameworks. While companies like Alibaba and Deepseek operate under China's rigorous internet regulations, their AI systems are designed to respect the directives that promote “basic socialist values”. This has implications for how models react to politically sensitive requests. For example, when he asked him questions about the status of Taiwan, QWQ-32B-PREVIEW provided an answer in accordance with the position of the Chinese government. Likewise, the prompts on Tiananmen square have led to non-responses, reflecting the regulatory environment in which these systems are developed.
While QWQ-32B-PREVIEW is marketed as available under authorized license, not all components of the model have been published. This partial opening limits the capacity to fully reproduce the model or to gain a complete understanding of its architecture. The debate on what constitutes an “opening” in the development of AI continues, with models ranging from entirely closed systems, offering only access to the API, to entirely open systems which disclose all the details, including weights and data. QWQ-32B-PREVIEW occupies common ground on this spectrum.
The rise of reasoning models such as QWQ-32B-PREVIEW arrives at a time when traditional “scaling laws” are called into question. For years, these laws suggest that the increase in data and IT resources would cause continuous improvement in AI capabilities. However, recent reports indicate that the rate of progress of the models of the main AI laboratories, in particular Openai, Google and Anthropic, began to set. This has stimulated a search for innovative approaches in the development of AI, including new architectures and techniques.
Such a winning approach to the field is the calculation of the test time, also called an inference calculation. This method allows AI models to use additional treatment time during tasks, improving their ability to manage complex challenges. The calculation of the test time forms the foundations of models such as O1 and QWQ-32B-PREVIEW, reflecting a change of orientation towards the optimization of performance during inference rather than relying on the training.
The main AI laboratories beyond Openai and Chinese companies also invest massively in reasoning models and a testing time calculation. A recent report pointed out that Google has considerably widened its team dedicated to reasoning models, which brought it to around 200 members. In addition to this expansion, the company has allocated substantial IT resources to advance this field of research on AI, signaling the growing commitment of industry in the future of AI reasoning.
