Microsoft AI announced on October 15 that it had completed pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and approximately 80 billion active per token. According to its technical report, training ran for 38 days on 16,000 NVIDIA GB200 GPUs, processing 50 trillion tokens.

A larger model with native vision

MAI-Base-2 integrates vision into pre-training without a separate image encoder. Vision-language data accounted for 18% of the training mixture. Microsoft says it processed all training data internally and used no distillation from other models.

The company also improved its MoE implementation and Muon optimizer. MoE-layer compute utilization rose from 24% to 35%, enabling expert layers throughout the model.

Strong internal results, with limits

Microsoft reports that MAI-Base-2 outperformed every open-weight model it tested on held-out evaluations. Its scaling-law analysis estimates 4.2–12.3 times greater training-compute efficiency than DeepSeek-V4-Pro-Base at equivalent evaluation scores.

That figure measures estimated pre-training efficiency—not inference speed, operating costs or a corresponding increase in intelligence. Some comparison models were also post-trained checkpoints, which Microsoft acknowledges can affect the comparison. Source: evaluation methodology

The next test is practical performance

MAI-Base-2 is undergoing post-training, and Microsoft plans to license weights to trusted partners through Foundry. Preparations for MAI-Base-3 are already underway.

The announcement strengthens Microsoft’s own model-development pipeline. The next question is straightforward: can this training progress translate into a model that reliably writes code, reasons through difficult problems and completes real work?

Sources & references