The Chinese company Alibaba has introduced an open-source artificial intelligence model called QwQ-32B-Preview. The model is capable of solving complex logical and mathematical problems, contains 32.5 billion parameters and can process queries up to 32,000 words in length. According to tests conducted by Alibaba, the QwQ-32B-Preview outperforms OpenAI’s o1-preview and o1-mini models in a number of areas.
Image source: Copilot
One of the key advantages of the QwQ-32B-Preview, as reported by TechCrunch, is its performance in the AIME and MATH tests. The AIME test evaluates the model’s performance against other artificial intelligences, and the MATH test is a set of word-based math problems. Thanks to its features, this AI bot is capable of solving complex problems and logic puzzles. However, it is noted that QwQ-32B-Preview is not without its drawbacks: it can unexpectedly change the language of communication, get hung up on certain points, and cope worse with tasks when common sense is required.
An interesting feature of the QwQ-32B-Preview is the ability to “check itself.” This helps avoid pitfalls common to other artificial intelligence systems, but slows down the process of finding solutions. Like OpenAI’s o1 models, Alibaba’s new development solves problems in stages, gradually analyzing and planning further steps.
Despite the fact that the QwQ-32B-Preview is available for download on the Hugging Face platform under the Apache 2.0 license (free software), which allows it to be used for commercial purposes, Alibaba has disclosed only part of the model’s components, thereby preventing us from studying the internal working mechanisms algorithm. Thus, despite statements about “openness,” the model remains an intermediate option between closed and open systems (open source license).
The Chinese origin of the model also imposes certain restrictions. Like other developments from China, QwQ-32B-Preview follows the requirements of a local regulator, which tests AI for compliance with “core socialist values.” The model will refuse to answer political questions related to China.
It is worth saying that interest in “reasoning” models is growing against the backdrop of a decline in the effectiveness of the traditional approach to increasing the capabilities of AI, based on simply scaling data and computing power. Companies such as OpenAI, Google and Anthropic are increasingly seeing their technology progress slow, which in turn is encouraging developers to look for new approaches, such as “test-time compute”, which give models additional time to process tasks. At the same time, large laboratories, including Google, are actively investing in the development of such technologies, which confirms their key role in the future of artificial intelligence.