The UAE launches low-cost AI inference model claiming to be 20 times more cost-effective than its competitors.

On Tuesday (September 9th) local time, the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in the UAE announced on its official website that it has jointly launched a low-cost inference model called "K2 Think" with G42. The press release claims that K2 Think only requires 32 billion parameters, yet surpasses inference models from other companies that are 20 times larger in scale. The model is based on the open-source Qwen 2.5 model from Alibaba and is run and tested on hardware provided by Cerebras. Hector Liu, the director of the MBZUAI Basic Model Research Institute, told the media that the team achieved high performance through various methods, including Chain of Thought Supervision Fine-tuning (CoT) - a progressive inference method, and test-time scaling - allocating more computing resources during the inference stage to improve performance. Liu stated, "What makes it special is that we treat it as a system, not just a model. Unlike ordinary open source models that are simply 'released', we actually deploy the model and observe how it improves over time." "It is hard to pinpoint one key step. It is more like a combination of a whole set of methods that leads to the final result." The press release stated, "K2 Think is not just a technological achievement, but a decisive moment in the field of artificial intelligence in the UAE." In recent years, the UAE has positioned itself as a global leader in AI to increase its geopolitical influence and reduce its reliance on crude oil. G42, an AI company in the UAE, is an example of this breakthrough. However, its competitors include neighboring Saudi Arabia, which established Humain through a sovereign fund in May, dedicated to building full-stack AI capabilities. From a broader perspective, the UAE's AI industry still has a long way to go to catch up with the US and China. The US's OpenAI and tech giants have taken the lead, and China has long made AI a strategic focus. Although the performance of K2 Think can be compared to OpenAI, developers emphasize that their goal is not to create another ChatGPT. Richard Morton, the general manager of the MBZUAI Basic Model Research Institute, explained that this model is designed for specific applications in mathematics, science, and other fields. He said, "The basic reasoning ability of the human brain is the cornerstone of all thinking processes. This application can significantly shorten the time it takes for a problem-solving or clinical experiment that usually requires 1000 to 2000 people and five years." Additionally, it allows regions lacking the capital and infrastructure like the US to use advanced AI technology. Morton concluded, "We are discovering that with fewer resources, we can actually achieve more." This article is reprinted from "Wealth Financial Network," edited by GMTEight: Liu Jiayin.