Microsoft finishes pretraining MAI-Base-2 artificial intelligence model

Microsoft has developed and completed pretraining on a new artificial intelligence base model named MAI-Base-2, according to reports from SiliconANGLE, Intelligent Living, XenoSpectrum, Breakingthenews.net, Tech Insider, and Technology Org. Outlets differ on the precise size of the model. SiliconANGLE reported that the model contains 2 trillion parameters, while Intelligent Living, XenoSpectrum, Tech Insider, and Technology Org put the figure at 2.8 trillion parameters. [1][2][3][4][5][6]

The model operates as a base prototype rather than a finished end-user model, SiliconANGLE reported. Base models require additional rounds of training by developers, which typically focus on adding reasoning features and safety guardrails, SiliconANGLE said. [1]

Microsoft claimed that the model offers major efficiency advantages over existing alternatives. Intelligent Living reported that the model uses a quarter of the compute, while XenoSpectrum and Tech Insider reported that Microsoft claims up to 12 times the training efficiency compared to DeepSeek. SiliconANGLE noted that DeepSeek-V4 is currently the most capable open-source base model on the market. [1][2][3][5]

As FTMQ Creators reported earlier, Microsoft AI Chief Executive Mustafa Suleyman told staff that an in-house model surpassed an OpenAI base model powering GPT 5.6 Sol on pre-training metrics using one-quarter of the compute power. [10]