Alibaba introduced QwQ-32B-Preview — a model for logical tasks

The Qwen team from Alibaba presented QwQ-32B-Preview — a new “logical” AI model, available for download under the open Apache 2.0 license. This model contains 32.5 billion parameters and can process context up to 32,000 words long. In some tests, such as AIME and MATH, it outperforms OpenAI o1-preview and o1-mini.

QwQ-32B-Preview is designed to solve complex logical problems and mathematical examples. According to the developers, its key advantage is the ability to “verify” its answers, improving accuracy. However, this approach requires more computation time. The model can also perform multi-step tasks, planning out solutions.

The model has certain limitations. It may switch between languages or exhibit cyclical reasoning, which can result in overly long answers. At the same time, QwQ-32B-Preview avoids discussions on political topics, which is typical for models from Chinese companies.

Although the model can be downloaded from the Hugging Face platform, some of its components remain unavailable. This limits the possibility of fully studying or reproducing QwQ-32B-Preview.