Microsoft AI completed pre-training of MAI-Base-2. It is an MoE model with 80 billion active and 2.8 trillion total parameters. MAI-Base-2 outperforms every open-weight model tested on held-out evaluations. It is 4-12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. Training used 16,000 GB200 GPUs over 38 days. Average MFU was 21% and goodput was 89%. Microsoft owns the full model pipeline end-to-end. It will license model weights to trusted partners via Foundry. MAI-Base-2 is now undergoing post-training. Microsoft used robust scaling ladders and model-infrastructure co-design. It also derived new scaling laws for hyperparameters. The model was trained on organic data with strong provenance. Microsoft plans to share more results soon.
What changed
MAI-Base-1 was a smaller, less efficient model with fewer parameters and no MoE architecture.
What it unlocks
Training a larger, more efficient model with improved reasoning performance and lower compute costs.
- 80B active parameters
- 2.8T total parameters
- 16k GB200 GPUs
- 38 days training time
Sources