Microsoft said it has completed pretraining for MAI-Base-2, its new in-house model. The system has 2.8 trillion total parameters and uses a mixture-of-experts architecture, with 80 billion parameters activated at a time.
According to the company, Microsoft used 16,000 NVIDIA GB200 GPUs across August and September this year. Training lasted 38 days and processed about 50 trillion tokens. MAI-Base-2 also introduces native vision pretraining for the first time, and visual data accounted for 18% of the training load.
Internal benchmark claims and compute comparison
Microsoft said internal tests showed MAI-Base-2 outperforming models including DeepSeek-V4-Pro-Base, Kimi K3 and Qwen3.8 on text prediction metrics in coding, question answering, mathematics, and science and engineering subjects.
The company also said that if the goal were only to train a model at the same level as DeepSeek-V4-Pro-Base, its method would require about one-quarter to one-twelfth of the compute. Microsoft said it continued to spend more compute to make the model larger and stronger, and that MAI-Base-2’s actual training compute still exceeded DeepSeek’s.
Mustafa Suleyman, Microsoft’s AI chief, said at an internal meeting on Oct. 15 that the pretrained base of MAI-Base-2 is already stronger than the OpenAI model supporting GPT-5.6 Sol, while using one-quarter of the training compute.
Training methods and rollout plan
Microsoft said it improved MoE expert routing, the Muon optimizer, and data filtering methods to reduce wasted memory and improve stability in large-scale training.
The company also said all training data was processed internally and that it did not obtain capabilities by distilling other companies’ models.
MAI-Base-2 is still in the post-training stage. Microsoft plans to use it in its own products first, then license model weights to enterprise partners through Foundry.