Alibaba Unveils 'Thinking' AI
Chinese company Alibaba has unveiled a “reasoning-focused” AI model, the QwQ-32B-Preview.
The proof-of-concept is presented by the AI-focused Qwen team. As part of the preliminary release, the neural network “demonstrates promising analytical capabilities” with some limitations.
QwQ-32B-Preview has 32.5 billion parameters, which reflects problem-solving skills.
Alibaba's AI outperforms OpenAI's o1 in both the AIME and MATH tests. The former evaluates the model's performance, while the latter consists of a set of word problems.
The QwQ-32B-Preview can solve logic puzzles and answer complex math questions. However, the Chinese tech giant noted that the model has problems such as mixing languages and switching between them unexpectedly, as well as falling into a cycle of circular thinking, which leads to long answers with no final result.
QwQ-32B-Preview and other reasoning-oriented models focus on efficient fact-checking, which results in more accurate answers. To do this, they require more time to think.
In October, it became known that Google was working on AI that could “reason like a human.”
To recap, in August, Alibaba launched a group of large, math-focused language models called Qwen2-Math that “outperform GPT-4o and Claude 3.5” in this area.