Alibaba’s newest AI model QwQ-32B-Preview outshines OpenAI’s o1 in some benchmarks

As competition intensifies in the AI field, Chinese retail giant Alibaba unveiled its QwQ-32B-Preview which reportedly outperforms OpenAI’s o1 series.

The latest model reportedly outshines OpenAI’s o1-preview and o1-mini models on some specific benchmarks, such as the AIME and MATH tests, which evaluate AI models’ performance in logic puzzles and math problems.

Alibaba has made QwQ-32B-Preview for download.

According to the retail giant, Alibaba’s new model is capable of tackling complex and intricate problems compared to normal large language models (LLMs) like ChatGPT-4 and Claude 3.5.

An article by Benzinga indicates that the QwQ-32B-Preview is one of the few available under a permissive license, enabling users to download and use it. The model is now available on the AI development platform Hugging Face.

However, Alibaba released certain components of the model to limit full replication of the model or insights into its working.

Alibaba’s latest model boasts 32.5 billion parameters which allows it to handle prompts of up to 32,000 words. With the model’s significant capabilities and semi-open accessibility, Alibaba’s new entrant sets the stage for a transformative leap in AI reasoning technologies.

Alibaba’s transparent announcement, which underscores its model’s sophistication, OpenAI has kept its parameter counts under wraps.

The coming of this model comes at a time when OpenAI is making significant strides in the AI sector. In October, OpenAI’s valuation jumped to $157 billion following a successful funding round.

Earlier this week, SoftBank Group SFTBF reportedly increased its stake in the ChatGPT maker through a $1.5 billion employee share buyout.

OpenAI is also said to be exploring the development of its own web browser to challenge Alphabet’s subsidiary Google Chrome browser after pressure coming from the US Department of Justice to divest it.

Alibaba admits the model has flaws too

Although it possesses some unique strengths, the new model also has limitations. According to the group, QwQ-32B-Preview has issues like unexpected language switches, which could potentially confuse users. The model also underperforms in tasks requiring common-sense reasoning, which is common with many AI systems.

According to AutoGPT, the model may get caught in logical loops, delaying responses.

Despite its shortcomings, its reasoning capabilities allow it to fact-check itself, therefore cutting on errors but increasing resolution time.

By reasoning through tasks and planning steps, Alibaba’s model avoids some pitfalls that affect traditional AI systems. But this approach demands extra time which might limit real-time application.

According to Benzinga, QwQ-32B-Preview’ responses align with Chinese regulatory standards, avoiding politically sensitive topics.

For example, politically sensitive topics like Taiwan will give responses that are aligned with the Chinese government’s stance.

Additionally, prompts about events like Tiananmen Square result in no responses coming up, showing the model’s cautious design.

While this might be ideal for the Chinese market, it can also limit its appeal on the global market. However, the model is a significant step into the world of reasoning AI. While its limitations may narrow its global appeal, its other components like logic and semi-open nature makes it a big competitor for OpenAI.

According to AutoGPT, QwQ-32B-Preview highlight the potential, and challenges, of this exciting frontier where AI labs across the world are working towards refining reasoning technology.

Source