Microsoft completes pre-training MAI-Base-2: 2.8T total and 80B active parameters, trained on 16,000 GB200 GPUs over 38 days; post-training is still underway.

Microsoft’s own held-out evaluations comparing MAI-Base-2 and other models; lower bits per byte is better. Figure 1 from its October 15 technical report.

Published 15 Oct 2026 · 16:32 UTC

Microsoft AI ↗ · Source dated 15 Oct 2026

Microsoft completes pre-training MAI-Base-2: 2.8T total and 80B active parameters, trained on 16,000 GB200 GPUs over 38 days; post-training is still underway. · ournews.ai