Microsoft AI finishes pre-training MAI-Base-2, a 2.8 trillion parameter MoE model
Microsoft AI says it has completed pre-training MAI-Base-2, a mixture-of-experts (MoE) foundation model with 80 billion active and 2.8 trillion total parameters. According to the company, the model outperforms every open-weight model it tested on its held-out evaluations.
- Microsoft AI completed pre-training of MAI-Base-2, an MoE model with 80 billion active and 2.8 trillion total parameters.
- The company says it outperforms every open-weight model it tested on held-out evaluations.
- Microsoft AI claims it is 4-12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base.
- The saving was put into training a larger model, according to the company.
- MAI-Base-2 follows MAI-Base-1 and the June MAI-Thinking-1 model.
What Microsoft AI announced
In a Lab Notes post dated October 15, 2026, Microsoft AI described MAI-Base-2 as its new state-of-the-art model. It uses an MoE design, meaning only part of the network is active for any given input. Microsoft AI puts the active share at 80 billion parameters, out of 2.8 trillion in total.
The post follows the company's June announcement of MAI-Thinking-1. That model is a general-purpose foundation model post-trained on top of MAI-Base-1, and Microsoft AI says it has strong reasoning performance. MAI-Base-2 is the next base model in this line.
Efficiency claims and comparisons
The headline claim concerns efficiency. Microsoft AI says MAI-Base-2 is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base. The company points to a Figure 3 in its post for this comparison. It adds that it put the saving into training a larger model, rather than simply spending less compute.
The company also reports that MAI-Base-2 outperforms every open-weight model it tested on its held-out evaluations. Its Figure 1 compares MAI-Base-2 with other models that have at least one trillion total parameters across a range of pre-training evaluations, where lower scores are better. These results come from Microsoft AI's own evaluations, and the source text gives no further detail on the individual benchmarks.
What it means
The post concerns pre-training only. MAI-Base-2 is a base model, the stage before post-training that turns a foundation model into a reasoning or chat product, as was done with MAI-Thinking-1 on top of MAI-Base-1. The source does not say whether a post-trained model built on MAI-Base-2 is planned, nor does it give a release date or availability.
What the announcement does show is Microsoft AI's focus on training efficiency at frontier scale. By its own account, it is converting compute savings into a larger model, and it measures itself against open-weight base models such as DeepSeek-V4-Pro-Base. Whether the claimed gains hold up will depend on how the model performs once it is post-trained and made available.
Microsoft AI: Frontier Pre-training at Microsoft AI | Microsoft AI
