A new type of artificial intelligence (AI) called QwQ-32B-Preview has recently been introduced. This AI model is designed to perform reasoning tasks and is one of the few that can compete with OpenAI’s model named o1. What makes QwQ-32B-Preview particularly interesting is that it is the first model of its kind that you can download and use freely under a permissive license.
QwQ-32B-Preview was created by a team at Alibaba called Qwen. This model is quite powerful, containing 32.5 billion parameters, which can be thought of as the building blocks that help the AI understand and solve problems. The model can handle prompts that are up to about 32,000 words long. In tests comparing its performance on various tasks, QwQ-32B-Preview has shown to be better than OpenAI’s earlier reasoning models called o1-preview and o1-mini. Generally, the more parameters an AI model has, the better it is at solving complex problems, although OpenAI does not share the number of parameters for its models.
According to Alibaba’s tests, QwQ-32B-Preview outperformed OpenAI’s o1-preview on two specific tests: AIME and MATH. The AIME test uses other AI models to assess performance, while the MATH test consists of various word problems that require mathematical reasoning. This new AI model is capable of solving logic puzzles and tackling challenging math problems because of its advanced reasoning abilities. However, it still has some limitations. For instance, it might unexpectedly switch languages, get stuck in repetitive loops, or struggle with tasks that require what we think of as “common sense.”
Read Also: Linkup Connects LLMs With Quality Content Sources (Legally)
One unique feature of QwQ-32B-Preview and similar reasoning models is that they can fact-check their own answers. This ability helps them avoid some common mistakes that other AI models make, although it does mean that they often take longer to find solutions. Like OpenAI’s o1, this new model approaches tasks by planning ahead and executing a series of steps to arrive at the correct answer.
You can download and run QwQ-32B-Preview from a platform called Hugging Face. It shares some similarities with another reasoning model called DeepSeek, especially in how it handles sensitive political topics. Since Alibaba and DeepSeek are both Chinese companies, they must comply with regulations set by China’s internet authorities. These regulations require that their models reflect the values promoted by the Chinese government. As a result, many Chinese AI systems avoid discussing controversial topics that could upset regulators. For example, when asked whether Taiwan is part of China, QwQ-32B-Preview confirmed that it is, aligning with the views of the Chinese government rather than the majority opinion worldwide. Inquiries about sensitive historical events, like the Tiananmen Square protests, were met with silence.
QwQ-32B-Preview is available for use under an Apache 2.0 license, which means it can be used in commercial projects. However, not all parts of the model have been shared publicly, making it difficult for others to replicate or fully understand how it works. The level of openness in AI models varies, with some being completely closed off (accessible only through an API) and others being more transparent (with their model architecture and data available). QwQ-32B-Preview falls somewhere in between.
The growing interest in reasoning models like QwQ-32B-Preview comes at a time when the traditional belief that simply increasing the amount of data and computing power for AI models would lead to better performance is being questioned. Recent reports suggest that major AI companies, including OpenAI and Google, are not seeing the same level of improvement in their models as they did in the past. This has led to a rush to explore new methods and technologies for AI development, including a technique known as test-time compute. This technique allows models extra processing time to complete tasks and is a key feature of both the o1 and QwQ-32B-Preview models.
In fact, other major companies, in addition to OpenAI and Chinese firms, are investing in the future of reasoning models through test-time compute. A recent report indicated that Google has significantly expanded its internal team focused on developing reasoning models, growing it to about 200 people and increasing the computing resources available to them.
