Microsoft AI disclosed it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, trained on 16,000 GB200 GPUs over 38 days in August and September 20261. The model outperforms every open-weight model Microsoft tested on held-out evaluations and is 4–12x more FLOPs-efficient than DeepSeek-V4-Pro-Base. MAI-Base-2 uses MoE in every layer across 102 layers and was pre-trained natively with vision using an encoder-free architecture. Microsoft said it will license model weights to trusted partners through Foundry and that post-training is underway.
Microsoft Completes MAI-Base-2, 2.8T-Parameter MoE Model Trained on 16k GB200 GPUs
Microsoft AI completed pre-training MAI-Base-2, an 80 billion active / 2.8 trillion total parameter MoE model trained on 16,000 GB200 GPUs over 38 days.