Alibaba recently introduced an inference-focused artificial intelligence (AI) model called Marco-O1. The model is similar to the QWQ-32B large language model, which is also optimized for tasks that require advanced inference functions, but an important difference is that MARCO-O1 is a smaller model and is extracted from the Qwen2-7B-Insructy model. The Chinese tech giant claims that several fine-tuning exercises have been used to make the new model reasoning focus on reasoning. Furthermore, the researchers stress that it has been optimized for complex real-world problem-solving tasks.
Alibaba Marco-O1 AI model
The new AI model is described in detail in a research paper published in the online preprint journal Arxiv. It is worth noting that papers published in online journals have not been peer-reviewed. Additionally, Alibaba hosts AI models to embrace the face and allows downloading and using it for personal and commercial use cases under the Apache 2.0 license.
However, since only partial datasets are provided, it is not fully open source. Therefore, users will not be able to copy or decompose models to analyze architectures or components.
To Marco-O1, it was fine-tuned according to the QWEN2-7B-Instruct Foundation model. In this article, the researchers emphasize that the AI model is powered by fine-tuning through thinking chains (COT), Monte Carlo tree search (MCT), reflection mechanisms, and other inference strategies.
As a result, Alibaba’s Marco-O1 can solve open-ended questions and find questions about “the lack of clear standards and rewards for the challenge.” However, it should be understood that advanced reasoning capabilities do not come from any advancement in hardware or architecture.
Instead, all inference models today use a technique called test time computing, which makes AI models spend more processing time on a single query. This allows them to test different theories to find solutions and fact checks. As a result, these models are designed to provide more accurate responses and accomplish complex tasks. According to the researchers, an important area where Marco-O1 excels in is understanding the nuances of spoken language and translated expressions.
According to the researchers, a limitation of the AI model claims that although Marco-O1 shows inference characteristics, “its performance is still not fully implemented.”
