Alibaba launches “reasoning” AI model

Chinese tech giant Alibaba has launched QwQ-32B-Preview, an advanced so-called “reasoning model” that stands as one of the few competitors to OpenAI’s o1 model. The model, which contains 32.5 billion parameters, can process up to 32,000-word instructions and outperforms OpenAI’s equivalent models in some benchmark tests, such as AIME and MATH.

Available through the AI platform Hugging Face under the Apache 2.0 license, QwQ-32B-Preview can be used for commercial purposes. Although the model is not fully open – some components have not been made public – it offers greater access than many other reasoning models on the market.

Reasoning models, such as QwQ-32B-Preview, differ from traditional language models in their ability to analyze and solve complex problems. While language models focus on generating and understanding text, reasoning models are designed to draw conclusions, solve logical problems, and review their own answers to ensure accuracy. This self-review makes them slower than language models but also more reliable on complex tasks.

QwQ-32B-Preview is specifically designed for solving logic and mathematical problems and includes mechanisms for reviewing and improving one’s own answers. However, the model is said to have some limitations, such as getting stuck in loops or underperforming on tasks that require common sense.