Alibaba has unveiled its open AI model, QwQ-32B-Preview.

The Chinese company Alibaba has announced the AI model QwQ-32B-Preview, which is capable of solving complex logical and mathematical problems. This system, featuring 32.5 billion parameters, can process queries up to 32,000 words long and demonstrates outstanding results in tests, reports TechCrunch.

According to TechCrunch, the model outperforms OpenAI counterparts, such as o1-preview and o1-mini, in several key metrics. Specifically, QwQ-32B-Preview has shown high performance in the AIME and MATH tests, which assess the AI's ability to solve text-based problems and interact with other systems.

One of the model's unique features is its ability for "self-checking," which helps avoid errors but slows down data processing. Similar to OpenAI systems, Alibaba's new model tackles tasks step-by-step, analyzing intermediate results.

Despite claims of openness, the model is only partially available. Its components can be downloaded from the Hugging Face platform under the Apache 2.0 license, allowing for commercial use, but the internal algorithms remain inaccessible. As a result, QwQ-32B-Preview occupies a middle ground between fully open and closed developments.

The model's Chinese origin has imposed certain limitations. It has been developed in accordance with local regulatory requirements and does not address politically sensitive questions related to China.

Experts note that the emergence of QwQ-32B-Preview highlights the importance of new approaches in AI development. Amid a slowdown in progress based on scaling computational power, developers are exploring concepts like "test-time compute." These technologies, actively supported by companies like Google, are shaping the future of artificial intelligence.