Alibaba's Qwen team has released QwQ-32 billion-Preview, a new AI reasoning model with 32.5 billion parameters.

QwQ-32 billion-Preview has self-fact-checking capabilities and can solve logic puzzles and math questions. The company claims that the model reasons through tasks by planning ahead and performing a series of actions to find answers, though it may switch languages unexpectedly, get stuck in loops, and underperform on common-sense reasoning tasks.

According to Alibaba's testing, QwQ-32 billion-Preview outperforms OpenAI's o1 models on the AIME and MATH tests. While the model is labeled as "open" under the Apache 2.0 license for commercial use, only certain components have been released, limiting the ability to replicate or fully understand the system.

Analyst QuickTake: Alibaba has previously released the Marco-o1 model, designed to advance open-ended reasoning in AI. The QwQ-32B-Preview appears to evolve this work, further advancing the reasoning capabilities. Additionally, the release of QwQ-32B-Preview signals Alibaba's ambitions to compete with OpenAI in the high-stakes field of reasoning AI. This development could spur further innovation and competition, benefiting the broader AI research community and end users.