Microsoft AI finishes pre-training MAI-Base-2
Microsoft AI says its new trillion-parameter-scale MoE base model beats every open-weight model it tested, which would give Microsoft a stronger in-house frontier model to license through Foundry.
Microsoft AI announced that it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. It succeeds MAI-Base-1, the base for MAI-Thinking-1, which Microsoft released in June. MAI-Base-2 is now being post-trained, and Microsoft has not given a release date.
Microsoft says MAI-Base-2 scored a lower bits-per-byte than every open-weight model it tested on held-out text. Its comparison set was DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. Only post-trained checkpoints of the last three were available, which Microsoft says may affect the results. The company also says its training recipe is 4.2–12.3x more FLOPs-efficient than DeepSeek-V4-Pro-Base's, based on scaling-law fits across four evaluations. Microsoft ran all of these evaluations itself. The code tests used its own codebase, and the STEM, math and Q&A sets were commissioned from vendors to limit contamination.
The run used 16,000 GB200 GPUs for 38 days in August and September 2026. Average model FLOPs utilization was 21% and goodput was 89%.
Microsoft credits three areas of work. The first is scaling ladders that use "source scaling" to account for data that repeats across epochs. The second is better MoE and Muon implementations along with activation gradient clipping. The third is expanded data pipelines, including vision data.
Microsoft says the model was trained on organic data with no distillation from other models. It says that data provenance will let it license the weights to trusted partners through Foundry.