Content
summary Summary

Alibaba has released QwQ-32B-Preview, a new AI model that focuses on logical reasoning and problem-solving capabilities. The model appears to match and sometimes outperform OpenAI's latest offerings in specific areas.

Ad

The Chinese tech giant's AI team, Qwen, says their new language model contains 32.5 billion parameters and can process up to 32,000 words of context. QwQ-32B-Preview shows particularly strong results in mathematical tests like AIME and MATH, with notable performance in the MATH-500 and GPQA benchmarks.

Comparison table: Performance benchmarks of six AI language models in four categories (GPQA, AIME, MATH-500, LiveCodeBench) with percentages.
QwQ matches and sometimes exceeds OpenAI's o1-preview in logic benchmarks. | Image: Qwen

Self-checking capabilities

Like OpenAI's o1 models, QwQ incorporates a self-verification system. It pre-plans its answers and double-checks its work, a process that adds to processing time but also boosts accuracy compared to typical language models. The Qwen team waxes philosophical about this feature:

QwQ embodies that ancient philosophical spirit: it knows that it knows nothing, and that’s precisely what drives its curiosity. Before settling on any answer, it turns inward, questioning its own assumptions, exploring different paths of thought, always seeking deeper truth. Yet, like all seekers of wisdom, QwQ has its limitations. This version is but an early step on a longer journey - a student still learning to walk the path of reasoning. Its thoughts sometimes wander, its answers aren’t always complete, and its wisdom is still growing. But isn’t that the beauty of true learning? To be both capable and humble, knowledgeable yet always questioning?

Qwen research team

The researchers acknowledge some shortcomings. QwQ can sometimes switch languages unexpectedly, get stuck in loops, and stumble over common-sense reasoning—common pitfalls for logic-focused language models.

Ad
Ad

Released under the Apache 2.0 license, QwQ is available for commercial use. However, Alibaba has only released certain components, making full replication impossible for now. A demo is available on Hugging Face.

Alibaba's cloud computing unit introduced the first Qwen models in August 2023. Qwen2, a more powerful successor, followed soon after, with improvements in programming, math, logic, and multilingual capabilities.

The current Qwen 2.5 series includes specialized versions: Qwen2.5 for general language, Qwen2.5-Coder for programming, and Qwen2.5-Math. Qwen2.5-Turbo, designed for larger context windows, was added recently.

China's Growing AI Presence

QwQ is the second "reasoning model" to come out of China. DeepSeek recently unveiled a similar system that also appears to challenge OpenAI's offerings. While both are currently only available as "mini" or preview versions, full releases could come later this year.

The arrival of these two Chinese models just weeks after OpenAI's o1 introduction raises interesting questions about OpenAI's competitive edge. However, the full capabilities of OpenAI's o1 model remain undisclosed, particularly regarding the potential of compute scaling. There might be more to these models than meets the eye, and architectural differences could still give OpenAI a distinct advantage.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Recommendation
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Alibaba has introduced QwQ-32B-Preview, a new AI model with 32.5 billion parameters that outperforms OpenAI's o1 models in certain benchmarks, particularly in logical reasoning and problem-solving tasks.
  • The model's self-checking ability, which involves planning answers in advance and verifying conclusions, enhances its accuracy but results in slower performance compared to traditional language models. Its limitations include language switching, loops, and tasks requiring 'common sense'.
  • Alibaba is not alone in this space; the Chinese company DeepSeek has also recently presented a "reasoning model" that rivals OpenAI's o1, which was introduced just two months ago.
Jonathan works as a freelance tech journalist for THE DECODER, focusing on AI tools and how GenAI can be used in everyday work.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.