Microsoft completes pretraining on new MAI-Base-2 large language model

Microsoft has developed a new large language base model called MAI-Base-2, according to reports from SiliconANGLE and Breakingthenews.net. The company announced the completion of its pretraining phase on October 15, Breakingthenews.net reported. SiliconANGLE reported that Microsoft detailed the underlying algorithm on Thursday. [1][4][5]

Outlets reported differing figures regarding the total parameter count of the model. SiliconANGLE reported that MAI-Base-2 has 2 trillion parameters. However, Intelligent Living, XenoSpectrum, and Tech Insider reported that the model features 2.8 trillion parameters. [1][2][3][5][6]

Microsoft claimed significant efficiency improvements during the training process. XenoSpectrum and Tech Insider reported that Microsoft claims up to a 12 times training efficiency gain over DeepSeek. Intelligent Living reported that the model required a quarter of the compute. [2][3][6]

MAI-Base-2 operates as a base model prototype rather than a finished LLM, SiliconANGLE reported. Developers transform base models into full products by conducting additional training rounds, which add reasoning capabilities and safety guardrails. [1][5]

Microsoft is an American multinational technology conglomerate headquartered in Redmond, Washington. The company has expanded its operations into cloud computing and artificial intelligence. [7]