微软公布2.8万亿参数MAI-Base-2模型
微软于10月15日宣布完成MAI-Base-2的预训练。这是一款总参数量2.8万亿、每次激活800亿参数的混合专家基础模型，使用1.6万颗Nvidia GB200 GPU训练38天，处理了包含文本和视觉数据的50万亿个token。微软称，该模型在其内部评估中超过了所测试的开放权重模型，并估算在选定测试中达到与DeepSeek-V4-Pro-Base相当的表现时，其训练方法的算力效率高4.2至12.3倍。微软尚未公布与所有主要专有模型的等效比较，其效率数据估算的是达到相当表现所需的算力，并不表示该模型普遍优于其他模型。
MAI-Base-2推动了微软扩大自研AI技术的努力。微软计划通过Microsoft Foundry向受信赖的合作伙伴授权模型权重，而非公开发布，以为企业客户提供更可预测的自定义AI部署路径。该模型仍处于后训练阶段，定价和普遍可用时间均未公布。
We may earn from vendors via affiliate links or sponsorships. This might affect product placement on our site, but not the content of our reviews. See our Terms of Use for details.
Microsoft is building a much larger AI model of its own, but the company says its latest breakthrough is about using computing power more efficiently.
The company on Oct. 15 announced that it has completed pre-training MAI-Base-2, a mixture-of-experts (MoE) foundation model with 2.8 trillion total parameters and 80 billion active parameters. The model succeeds MAI-Base-1, which Microsoft introduced earlier this year.
According to Microsoft’s technical announcement, MAI-Base-2 outperformed the open-weight models included in its internal evaluations. The company estimates its training approach is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base when matching performance on selected tests.
Microsoft trained the model using 16,000 Nvidia GB200 GPUs over 38 days in August and September. Its training run processed 50 trillion tokens, including text and visual data.
MAI-Base-2 activates only a fraction of its parameters for each input, allowing it to retain a much larger overall model without using every parameter for every operation.
#Microsoft#MAI-Base-2#Microsoft Foundry#Nvidia GB200#DeepSeek-V4-Pro-Base#OpenAI