Microsoft completes pre-training of a 2.8-trillion-parameter model, outperforming the GPT-5.6 base model with only one-quarter of the computing power.

Beating AI News: Microsoft announces completion of pre-training for its new self-developed model MAI-Base-2. The model has a total of 2.8 trillion parameters, adopts a Mixture-of-Experts (MoE) architecture, and activates only 80 billion parameters per forward pass. Between August and September this year, Microsoft used 16,000 NVIDIA GB200 GPUs for 38 days of training, processing approximately 50 trillion tokens. The new model also introduces native visual pre-training for the first time, with visual data accounting for 18% of its training data. Internal tests released by Microsoft show that MAI-Base-2 outperforms models including DeepSeek-V4-Pro-Base, Kimi K3, and Qwen3.8 on text prediction benchmarks for code, question answering, mathematics, and science and engineering disciplines. Microsoft also claims that training a model of equivalent performance to DeepSeek-V4-Pro-Base would require only 1/4 to 1/12 of the computing power using its approach. However, Microsoft has invested additional computing power to build a larger, more capable model; MAI-Base-2’s actual training computing power consumption is still higher than that of DeepSeek. Microsoft AI head Mustafa Suleyman stated at an internal meeting on October 15 that MAI-Base-2’s pre-training base is already stronger than the OpenAI model supporting GPT-5.6 Sol, while requiring only one-fourth the training computing power. Microsoft improved MoE expert scheduling, Muon optimizer, and data filtering methods to reduce GPU memory waste and enhance the stability of large-scale training. The company also emphasized that all training data is processed in-house, and the model’s capabilities were not obtained through distillation of other firms’ models. MAI-Base-2 is currently in the post-training phase. Microsoft plans to first deploy it for its own products, then license the model weights to enterprise partners via Foundry.