MAI-Base-2 Explained: Microsoft's 2.8T-Parameter Base Model

Microsoft AI’s MAI-Base-2 is a pre-trained mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters per token, trained natively on text, code, images, and 50 trillion tokens using 16,000 GB200 GPUs over 38 days. Microsoft reports that its training recipe is 4.2–12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base and that MAI-Base-2 achieved lower bits-per-byte scores on its internally held-out code, QA, STEM, and math datasets than the open-weight models it compared, although these results cannot be independently reproduced and measure text prediction rather than agent performance. The model remains in post-training, has no public release date or endpoint, and Microsoft intends to license weights to trusted Foundry partners; MAI-Thinking-1, based on the earlier MAI-Base-1 model, is currently available in preview. Architectural and training changes from MAI-Base-1 include MoE layers throughout the network, a revised routing design, native vision capabilities, larger and more diverse data, Muon optimization, source-scaled experiments, and activation-gradient clipping that also identified a data-processing issue. The discussion emphasizes that organizations considering future MAI models should evaluate them against their own held-out prompts and production requirements, including instruction following, tool use, safety, grounding, long-context reliability, and consistency, rather than relying solely on pre-training metrics such as bits per byte.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.