Microsoft AI Completes Pre-training of MAI-Base-2, a 2.8 Trillion Parameter MoE Model

Microsoft AI Completes MAI-Base-2 Pre-training

Microsoft AI has completed pre-training of MAI-Base-2, which the company describes as a new state-of-the-art foundation model. The model uses a mixture-of-experts architecture with 80 billion active parameters and 2.8 trillion total parameters, placing it among the largest pre-trained base models in the company's portfolio.

The work follows the June announcement of MAI-Thinking-1, a general-purpose foundation model post-trained on top of MAI-Base-1 with strong reasoning performance. MAI-Base-2 builds directly on that lineage as the next-generation base model.

Outperforming Open-Weight Rivals

According to the company's Lab Notes blog, MAI-Base-2 outperforms every open-weight model tested on Microsoft's held-out evaluations. The internal benchmarks cover code, question answering, math, and STEM evaluation sets, with performance measured in held-out bits per byte where lower scores are better.

Gains in Pre-training Efficiency

The team reports that MAI-Base-2 is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which it identifies as the strongest open base model in the comparison. Microsoft says it redirected the compute savings from that efficiency into training a larger model, a trade-off it presents as a deliberate design choice for frontier pre-training.

The company's comparisons also reference models with at least one trillion total parameters, including post-trained systems such as MiMo-V2.6-Pro-RL, Qwen3.8-2.4T-A95B, and Kimi K3.

Sources: Microsoft AI · Techmeme