According to CoinMeta, Microsoft has announced the completion of pre-training for its new generation of self-developed model mai-base-2, which boasts a total of 2.8 trillion parameters. This model utilizes a moe hybrid expert architecture, activating 80 billion parameters each time. From August to September this year, Microsoft employed 16,000 NVIDIA GB200 GPU units to train the model for 38 days, processing approximately 50 trillion token. The new model introduces native visual pre-training for the first time, with visual data accounting for 18% of the total training volume. Internal tests have shown that mai-base-2 outperforms models such as deepseek-v4-pro-base, kimi, k3, and qwen3.8 in multiple text prediction indicators. Microsoft claims that to train a model of equivalent level to deepseek-v4-pro-base, only 1/4 to 1/12 of the computational power would be required. The person in charge of AI, Mustafa Suleyman, stated that the pre-training foundation of mai-base-2 is already stronger than that of the OpenAI model which supports GPT-5.6, and the training computational power needed for mai-base-2 is merely one-fourth of the latter's. This model is still in the post-training phase; Microsoft plans to use it in its own products and will license the model weights to corporate partners through Foundry.