Recently, Alibaba's Qwen team unveiled QwQ-32B-Preview, a next-generation inference-based artificial intelligence model. This model boasts 32.5 billion parameters and is capable of handling input prompts up to 32,000 words in length. In several benchmark evaluations, it outperformed OpenAI's o1-preview and o1-mini inference models. Although the number of parameters typically correlates with a model's problem-solving abilities, OpenAI has not disclosed the exact parameter count of their models.
Testing revealed that QwQ-32B-Preview excels in AIME and MATH assessments, which evaluate performance through other AI models and the ability to solve textual problems, respectively. The model is proficient in handling logic puzzles and addressing moderately challenging mathematical questions. However, it has certain limitations, such as occasionally switching languages unexpectedly, entering into loops, or underperforming in tasks that require common-sense reasoning.
Notably, QwQ-32B-Preview features self-fact-checking capabilities that help minimize errors, though this also makes the problem-solving process more time-consuming. Similar to OpenAI's o1, QwQ-32B-Preview resolves issues by pre-planning and executing a series of actions.
QwQ-32B-Preview is available for download and operation on the Hugging Face platform under the Apache 2.0 license, making it suitable for commercial use. However, only select components are publicly accessible.