The Chinese company Alibaba has introduced the QwQ-32B-Preview, an AI model “focused on reasoning capabilities.”
This experimental development is presented by the AI-focused Qwen team. In its preliminary release, the neural network “demonstrates promising analytical abilities” albeit with some limitations.
The QwQ-32B-Preview boasts 32.5 billion parameters, a figure indicative of its problem-solving skills.
Alibaba’s AI surpasses OpenAI’s o1 in the AIME and MATH tests. The former evaluates model performance, while the latter comprises a set of verbal tasks.
The QwQ-32B-Preview can solve logical puzzles and answer complex mathematical questions. However, the Chinese tech giant noted issues such as language mixing and unexpected switching between languages, as well as falling into circular reasoning, resulting in lengthy answers without a definitive conclusion.
The QwQ-32B-Preview and other reasoning-focused models are concentrated on effective fact-checking, which allows for more accurate responses. This requires them to take more time for contemplation.
In October, it was reported that Google is working on AI “capable of reasoning like a human.”
Back in August, Alibaba launched a group of large language models focused on mathematics under the name Qwen2-Math, which “surpass GPT-4o and Claude 3.5” in this field.