Microsoft says its new MAI-Base-2 model beats open rivals with less compute

Microsoft has unveiled MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters per token. Microsoft AI announces in an official blog post that the model is now moving into post-training, the stage where developers adapt a base model for tasks such as chat, reasoning, and instruction following.

The company says MAI-Base-2 performs better than every open-weight model it tested on its internal held-out evaluations. Microsoft also claims its training recipe is 4.2 to 12.3 times more efficient in pre-training FLOPs than DeepSeek-V4-Pro-Base, a competing open base model. FLOPs are a measure of the computing work used to train AI systems.

MAI-Base-2 uses a mixture-of-experts, or MoE, design. This approach contains many specialized parts, called experts, but activates only a subset for each piece of input. Microsoft says this allows the model to grow significantly while limiting the amount of computing needed for each response.

Large-scale training and data

Microsoft trained the model on 16,000 Nvidia GB200 GPUs for 38 days. The training mix contained 50 trillion tokens, including text, code, and visual material. Vision data accounted for 18 percent of the mixture, which means the model was trained from the start to process images alongside text.

The company says it expanded its corpus by about 36 percent compared with MAI-Base-1. It emphasizes that its data comes from internally processed “organic” sources rather than output distilled from other AI models. Microsoft says its web crawler follows robots.txt and related website controls.

Microsoft attributes the reported efficiency gains to data quality, new hyperparameter scaling methods, and infrastructure changes. These include a revised MoE system designed to avoid dropping tokens when individual experts receive heavy demand, plus optimizations to the Muon training optimizer.

Microsoft has not released model weights or post-training benchmarks. It says it plans to license weights to trusted partners through its Foundry platform after further work.