Details
- Microsoft AI says it has completed pretraining MAI-Base-2, an in-house foundation model in the roughly 3-trillion-parameter class.
- The model uses a mixture-of-experts design, activating only a portion of its parameters for each task to improve computational efficiency.
- On held-out code evaluations, Microsoft says MAI-Base-2 matched DeepSeek-V4-Pro-Base’s pretrained score using 12 times less training compute.
- The company attributes the clean training run to scaling ladders, iterative recipe refinement, and improved efficiency as model size increased.
- MAI-Base-2 remains in post-training, with additional results expected before any broader product or model release details.
- The announcement marks a larger in-house model effort for Microsoft AI, following earlier MAI models and expanding its role beyond relying primarily on external model providers.
- Microsoft’s technical material presents the result as a pretraining milestone rather than a finished assistant or commercially available API.
Impact
MAI-Base-2 strengthens Microsoft’s position as it builds proprietary models alongside partnerships with leading external developers. The claimed compute advantage over DeepSeek-V4-Pro-Base could lower the cost of training similarly capable systems, although the comparison currently covers pretraining evaluations rather than post-trained user performance. Its practical significance will depend on forthcoming results, deployment plans, inference costs, and whether Microsoft releases access through its enterprise platforms.