IT之家 10 月 16 日消息,Microsoft(微软)当地时间昨日宣布,其已经完成了新一代 SOTA 基础模型 MAI-Base-2 的预训练工作。这一模型在表现维度强于所有参测开源权重模型。
MAI-Base-2 总参数规模达 2.8T,活跃参数规模 80B,在 6,000 个 NVIDIA GB200 GPU 上历经 38 天训练而成。在整个训练过程中,该模型平均浮点算力利用率为 21%、有效吞吐量为 89%。
依据微软的说法,MAI-Base-2 的预训练算力效率是 DeepSeek-V4-Pro-Base 的 4~12 倍。
微软表示,三项举措对于 MAI-Base-2 的成功预训练尤为关键:科学规模化、模型-基础设施协同外观、数据规模化。在完善 MAI-Base-2 的同时,这家科技巨头已启动 MAI-Base-3 的准备工作。
相关阅读:
参考
本文来源:https://www.ithome.com/1/013/835.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT House October 16th news, Microsoft (Microsoft) local time yesterday announced that it has completed the pre-training of the new generation Sota basic model MAI-Base-2. This model is stronger in the performance dimension than all the open source weight models measured.
MAI-Base-2 has a total parameter scale of 2.8T, an active parameter scale of 80b, and 38 days of training on 6,000 NVIDIA GB200 GPUs. Throughout the training process, the model had an average floating-point hashrate utilization of 21% and an effective throughput of 89%.
According to Microsoft, MAI-Base-2's pre-trained budget is 4-12 times more efficient than DeepSeek-V4-Pro-Base.

Microsoft said that three initiatives are particularly critical to the successful pre-training of MAI-Base-2: scientific scale, model-infrastructure collaborative appearance, and data scale. While perfecting MAI-Base-2, the tech giant has launched preparations for MAI-Base-3.
Read Next:
REFERENCE
This article is from: https://www.ithome.com/1/013/835.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.