Microsoft AI finishes pretraining MAI-Base-2, a 2.8T-parameter MoE model
What happened
On October 16, Microsoft AI published pretraining details for MAI-Base-2 and said pretraining is complete. The model is a mixture-of-experts base model with 2.8 trillion total parameters and 80 billion active parameters. Reported compute-efficiency comparisons put its FLOPs efficiency at 4.2 to 12.3 times that of DeepSeek-V4-Pro-Base. In Microsoft AI's own evaluations, it beat every open-weight model tested. What has been disclosed so far covers the finished pretraining run, model size, compute efficiency and those in-house results.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- AI HOT · ModelsMicrosoft AI 公布 MAI-Base-2 预训练细节：2.8T 参数 MoE 模型 FLOPs 效率达 DeepSeek-V4-Pro-Base 的 4.2 至 12.3 倍
Microsoft AI 完成 MAI-Base-2 预训练，这是一个 80B 激活、2.8T 总参数的 MoE 基础模型，在自建评测上优于所有测试过的开放权重模型。
Heat over time
Not enough continuous observations to draw a trend yet.