Alibaba unveiled the OpenAI o1 competitor model with reasoning ability; In mathematical benchmarks, this model has performed better than o1; But for now it is available as a test.
The Chinese company Alibaba, which is the world’s largest retailer and one of the largest e-commerce companies in the world, has unveiled an artificial ielligence model capable of reasoning, which is considered a new competitor to OpenAI’s o1 model.
The iroduced model coains 32.5 billion parameters and can respond to requests with a maximum of 32 thousand tokens.
Like other large models, this model’s performance is reasoning, in the sense that during its inference, the AI uses more computing cycles to check the answers it was to provide to the user and correct mistakes.
As a result of this capability, this model is better suited for tasks that require logical reasoning and programming, such as math and coding.
This model is called QwQ and it was able to defeat o1-preview in AIME and MATH benchmarks that evaluate the model’s ability to solve mathematical problems.
This model also performed better than o1-mini in the GPQA benchmark, which evaluates scieific reasoning; But in the field of coding and based on the LiveCodeBench benchmark, the o1 model has performed better; It should be said that the performance of QwQ was better than other models such as GPT-4o and Claude 3.5 Sonnet.
Alibaba’s AI model is currely available as a preview, and we can expect more improved models from the company in the future.
“Through our deep explorations and couless experimes, we discovered something very tangible: when we take the time to think, question, and reflect, the model’s understanding of mathematics and programming blossoms like a flower in the sun,” Alibaba said in a stateme. The process of careful reflection and irospection leads to significa improvemes in solving complex problems.
However, the company didn’t say anything about the data or the process it we through to train its model, but given that QwQ is an open-source model, its thought process is not hidden, and it is possible to understand how the model reasons when solving problems. , we to its text.
