Alibaba Unveils a Competitor to OpenAI’s o1 with a Twist

Alibaba has just revealed a new AI model, QwQ-32B-Preview, that's ready to challenge OpenAI's o1 in the "reasoning" AI race. Developed by Alibaba's Qwen team, this model boasts 32. 5 billion parameters and can handle prompts up to 32, 000 words long. It outperforms OpenAI's o1-preview and o1-mini on certain benchmarks like AIME and MATH. QwQ-32B-Preview can solve logic puzzles and tough math questions, but it has its faults. It might switch languages randomly, get stuck in loops, and struggle with tasks that need "common sense reasoning. " What sets QwQ-32B-Preview apart is its ability to fact-check itself. This means it takes longer to arrive at solutions but avoids common pitfalls. The model isn't without its political sides, though. As a Chinese company, Alibaba has to play by the rules set by China's internet regulator. This means the model might not touch on sensitive topics like speculation about the Xi Jinping regime.

QwQ-32B-Preview is "openly" available under an Apache 2. 0 license, which allows for commercial use. However, not all parts of the model have been released, so replicating it or understanding its inner workings is a challenge. The debate around the "openness" of AI models continues, with QwQ-32B-Preview falling somewhere in the middle. The recent focus on reasoning models comes as the effectiveness of "scaling laws" is under scrutiny. These laws suggest that more data and computing power will continually boost a model's capabilities, but recent findings show that this might not be the case. This has prompted a search for new AI approaches, including test-time compute, which gives models extra processing time. Companies like Google are also investing heavily in this area.

Actions