Microsoft MAI-Base-2 预训练完成，FLOPs 效率达 DeepSeek-V4-Pro-Base 的 4-12 倍
微软自家基座模型 MAI-Base-2 训完了，号称算力效率是 DeepSeek-V4-Pro-Base 的 4-12 倍，还顺手把省下的算力拿去训更大的模型。
Microsoft 宣布 MAI-Base-2 完成预训练，MAI-Base-3 训练周期已在准备中。据官方说法，MAI-Base-2 在其内部 held-out 评估中优于所有受测的开放权重模型。在预训练 FLOPs 效率上，它比最强的开源基座模型 DeepSeek-V4-Pro-Base 高 4-12 倍，节省的算力被用于训练更大的模型。Figure 1 对比了 MAI-Base-2 与其他总参数量至少 1 万亿的模型在多组预训练评估上的表现。
MICROSOFT 🔥: MAI-Base-2 pre-training has been completed; MAI-Base-3 training cycle is already being prepared as well.
> "MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations. It is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base (see Figure 3 below), and we have put that saving into training a larger model."
* Figure 1 shows the performance of MAI-Base-2 and other models with at least one trillion total parameters across a range of pre-training evaluations (lower is better).
