Microsoft reportedly completes MAI-Base-2 pretraining, with a parameter scale of 2800000000000

Article

Microsoft completes 2800000000000 parameter-model pretraining: surpassing GPT-5.6 the base model, using only 1/4 Beating AI News Flash reports that Microsoft has announced completion of pretraining for its next-generation self-developed model MAI-Base-2. The model has a total of 2800000000000, uses a Mixture-of-Experts (MoE) architecture, and activates only 80000000000 parameters each time.

From 8 to September, Microsoft deployed 16000 cards NVIDIA GB200 GPU, trained for 38 days, and processed approximately 50000000000000 Token. The new model also introduced native vision pretraining for the first time, with visual data accounting for 18%of the training volume.

Microsoft's internal tests show that MAI-Base-2 surpassed models such as DeepSeek-V4-Pro-Base、Kimi K3、Qwen3.8 on text-prediction metrics in coding, question answering, mathematics, and science and engineering disciplines.

Microsoft also claimed that, if the goal is only to train a model at the same level as DeepSeek-V4-Pro-Base, its method requires approximately 1/4 to 1/12 of the computing power. However, Microsoft continued investing more compute to make the model larger and stronger. The MAI-Base-2 model's actual training compute consumption was still higher than DeepSeek's.

Microsoft AI head Mustafa Suleyman said at an internal meeting of October 15 that the pretraining base of MAI-Base-2 was already stronger than the OpenAI model supporting GPT-5.6 Sol, while using only one-quarter of the latter's training compute.

Microsoft improved MoE expert routing, the Muon optimizer, and data-filtering methods, reducing GPU-memory waste and improving the stability of large-scale training. Microsoft also emphasized that all training data was processed in-house and that it did not acquire capabilities by distilling other companies' models.MAI-Base-2 is currently still in the post-training stage. Microsoft plans to use it first in its own products, then license the model weights to enterprise partners through Foundry.

Original link https://m.theblockbeats.info/flash/372232

Key points

Microsoft announces completion of pretraining for its self-developed model MAI-Base-2.
MAI-Base-2 Total parameters reach 2800000000000, using an MoE architecture and activating 80000000000 parameters each time.
Microsoft used 16000 cards NVIDIA GB200 GPU for training over 38 days, processing approximately 50000000000000 Token。
Visual data accounted for 18%of the training volume, and the model is currently still in the post-training stage.
Microsoft said MAI-Base-2 surpassed DeepSeek-V4-Pro-Base、Kimi K3 and Qwen3.8。
Microsoft said its training method requires approximately one-quarter to DeepSeek-V4-Pro-Base of the compute required by relevant methods to train a model at the same level as 12-th, but MAI-Base-2 actual training compute consumption was still higher than DeepSeek's.
Mustafa Suleyman said that MAI-Base-2 the pretraining base was stronger than the OpenAI model supporting GPT-5.6 Sol, with training compute equal to one-quarter of the latter's.
Microsoft said it improved MoE expert routing, the Muon optimizer, and data-filtering methods.
Microsoft plans to use MAI-Base-2 first in its own products, then license the model weights to enterprise partners through Foundry.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a large model that uses a mixture-of-experts architecture: although it has many total parameters, only a subset of them is activated at any given time. Microsoft also claims that it performs better on multiple text-based tasks and reduces wasted computational resources through its training methods. However, these performance and efficiency comparisons currently come mainly from Microsoft's internal testing and statements by related personnel.

Why it matters to readers

If the relevant claims are substantiated, Microsoft may strengthen its self-developed foundation-model capabilities and advance its product and Foundry model-licensing plans.

It demonstrates that large models may improve efficiency not only by increasing the number of parameters, but also by activating only a subset of parameters and improving training methods.

Risks and unknowns

The provided materials do not include public benchmark tests, a technical report, or original conference records.
The comparisons of performance, computational resources, and model capabilities in the article are based mainly on Microsoft's internal testing and the framing used in conference remarks.
It is not yet clear MAI-Base-2 when post-training will be completed and integrated into the product or Foundry.

Microsoft reportedly completes MAI-Base-2 pretraining, with a parameter scale of 2800000000000

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center