Alibaba Launches Open Source Reasoning Model to Rival OpenAI's O1
Alibaba introduces QwQ-32B-Preview, an open-source reasoning AI that outperforms OpenAI's o1 on key math benchmarks. The 32.5-billion-parameter model self-fact-checks but exhibits Chinese political censorship.
Alibaba releases QwQ-32B-Preview, a new open-source reasoning AI model that directly challenges OpenAI's o1. The model contains 32.5 billion parameters, handles prompts up to 32,000 words, and outperforms OpenAI's o1-preview and o1-mini on rigorous math benchmarks like AIME and MATH.
Like other reasoning models, QwQ-32B-Preview effectively fact-checks itself by planning ahead and working through a series of logical actions to find solutions. However, this self-correction process means the model takes longer to generate answers, and it still struggles with issues like getting stuck in loops or failing at common sense reasoning tasks.
Developers can download and run the model from Hugging Face, but it comes with the political limitations expected from a Chinese tech company. The AI avoids sensitive topics and echoes the Chinese government's stance on issues like Taiwan, reflecting the regulatory requirements imposed by China's internet authorities.