Microsoft AI completes pre-training of MAI-Base-2, claims 4-12x efficiency over DeepSeek-V4-Pro-Base
Microsoft AI announced the completion of pre-training for MAI-Base-2, a 2.8T parameter MoE model, claiming its pre-training compute efficiency is 4 to 12 times higher than DeepSeek-V4-Pro-Base.
ImportanceStructuralEvidenceE2 unreplicatedWrite-upQuick
Microsoft AI announced on October 15, 2026, in its official 'Lab Notes' blog that it has completed the pre-training of its new foundation model, MAI-Base-2. The model utilizes a Mixture-of-Experts (MoE) architecture with 2.8 trillion total parameters and 80 billion active parameters.
In the announcement, Microsoft stated that MAI-Base-2 outperforms every open-weight model tested on their held-out evaluations. Specifically, Microsoft claims that its pre-training FLOPs efficiency is 4 to 12 times higher than that of DeepSeek-V4-Pro-Base, currently the strongest open base model.
This release covers only the pre-trained base model. Microsoft indicated that post-training results, including reasoning performance, will be shared soon. The efficiency comparisons are based on Microsoft's internal testing and have not been independently verified or accompanied by public benchmark details.