Microsoft Completes Pre-Training of 2.8-Trillion-Parameter MAI-Base-2 Model
What happened
On October 15, 2026, Microsoft announced the completion of pre-training for its MAI-Base-2 model, a mixture-of-experts foundation model with 2.8 trillion parameters.
The Context
- Innovative architecture: MAI-Base-2 utilizes a mixture-of-experts design, activating 80 billion parameters per token, enhancing efficiency and performance.
- Competitive landscape: This development comes amid rising competition in AI pre-training, with Microsoft focusing on proprietary data pipelines and integration of vision data.
- Future licensing: Microsoft plans to license the model's weights to select partners via Azure AI Foundry, potentially reshaping the competitive dynamics in AI.
The Number
— the total parameters in the MAI-Base-2 model, which signifies a leap in AI capabilities and efficiency, crucial for professionals relying on advanced AI solutions.
Takeaway
As Microsoft continues post-training, expect significant advancements in proprietary AI applications that could redefine industry standards.
AI news with an enterprise and cloud focus.
"Covers AI in the context of data infrastructure, cloud, and enterprise stacks."
— A47 Editor
Microsoft reveals cutting-edge MAI-Base-2 model with 2T parameters
Microsoft Corp. has unveiled its new MAI-Base-2 model, a cutting-edge artificial intelligence base model featuring 2 trillion parameters, which is designed to be more efficient than existing open-source alternatives. This announcement was made by the...
News for senior developers on AI/ML and data engineering.
"Conference-linked outlet for practitioner news and Q&As."
— A47 Editor
Microsoft Commits Maia to Multi-Gigawatt Inference, and Its Efficiency Metric Draws Questions
Microsoft has announced plans to scale its Maia accelerator family to multi-gigawatt deployments starting in 2027, positioning the Maia 300 as the primary inference system for its products. This new architecture integrates scale-up and scale-out capa...
Curated tech headlines including AI stories.
"Influential aggregator surfacing the day’s top tech/AI links."
— A47 Editor
Microsoft unveils MAI-Base-2, a MoE model with 2.8T total and 80B active parameters, claiming 4x-12x more pretraining FLOPs-efficiency than DeepSeek-V4-Pro-Base (Microsoft AI)
Microsoft has introduced MAI-Base-2, a new mixture of experts (MoE) model featuring a total of 2.8 trillion parameters and 80 billion active parameters, claiming to achieve 4x-12x greater pretraining efficiency in FLOPs compared to its predecessor, D...
Business and tech news excluding paywalled content.
"High-volume business/tech outlet with frequent AI coverage."
— A47 Editor
Microsoft AI chief says the company beat OpenAI at its own game, using a quarter of the computing power
Microsoft AI CEO announced that the company's latest AI model has surpassed an OpenAI base model in specific pre-training metrics while utilizing only a quarter of the computing power. This achievement highlights Microsoft's advancements in AI techno...