Microsoft AI unveiled MAI-Base-2 on October 15, a 2.8-trillion-parameter mixture-of-experts model that has completed pre-training but is not yet available to try. The company says it trained the model in 38 days using 16,000 Nvidia GB200 GPUs, with post-training now under way.
The announcement focuses on the model’s architecture, training process and text-prediction performance rather than its capabilities as a finished assistant. Microsoft has not announced a release date, pricing or downloadable weights. It describes plans to license weights to trusted partners through Microsoft Foundry, rather than an open-weight release.
Developed by the Microsoft AI unit led by Mustafa Suleyman, MAI-Base-2 activates 80 billion parameters for each token despite having 2.8 trillion parameters overall. Its routing system selects seven of 512 experts, alongside one shared expert that remains active. This sparse architecture allows a model to increase its total parameter count without using the entire network for every token.
The new model is substantially larger than MAI-Base-1, which had 1 trillion total parameters and 35 billion active parameters. Training increased from 30 trillion to 50 trillion tokens, while the network grew from 78 to 102 layers. Unlike its text-only predecessor, MAI-Base-2 incorporated visual data from the start of pre-training.
Microsoft’s central performance claim concerns bits per byte, or BPB, a measure of how efficiently a model predicts previously unseen text. Lower scores indicate better prediction. Because the measurement is expressed in bytes, it supports comparisons between models that divide text into different tokens.
Microsoft says MAI-Base-2 achieved lower BPB than the open-weight models in its comparison. However, the evaluations used internal datasets, and some competing checkpoints had already undergone post-training, which can change their results. The comparison also excluded proprietary models whose base checkpoints are not publicly available. Those limits make the findings narrower than a general claim of leadership across AI tasks.
The company also estimates that its training recipe would require 4.2 to 12.3 times less computation to match the BPB of DeepSeek-V4-Pro-Base. That estimate comes from scaling-law projections using smaller Microsoft models, not a controlled comparison in which both systems received equal training budgets. Microsoft states that MAI-Base-2 itself consumed more training computation than the DeepSeek model, with the claimed efficiency gains used to support greater scale.
Several engineering changes underpin the training approach. Microsoft used the Muon optimizer for weight matrices outside the embeddings and placed mixture-of-experts components in every layer, replacing its predecessor’s alternating dense and expert layers. It also introduced routing designed to retain tokens when experts become overloaded, alongside changes to gradient clipping intended to limit the disruption caused by anomalous training examples.
Data provenance is another part of the announcement. Microsoft says it processed the training corpus internally and did not use distillation from other models. The data has a March 2026 knowledge cutoff, with code accounting for 45% of the training mix. Visual material includes captioned images and documents combining text and figures.
The remaining milestone is post-training, which adapts a base model for instruction-following and other practical uses. Microsoft has not yet released MAI-Base-2 results for coding, agents, mathematics or computer-use tasks. Those evaluations, along with access terms and a Foundry availability date, remain outstanding.
Sources: Pasquale Pillitteri