Microsoft reveals cutting-edge MAI-Base-2 model with 2T parameters
SiliconANGLE Maria Deutscher ● Covered by 3 sources
Microsoft built a new base AI model called MAI-Base-2. It’s said to beat DeepSeek-V4-Pro-Base while using far less hardware.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft says MAI-Base-2 is a fresh take on a base model, not a finished chatbot. That distinction matters. Base models are the rough core that gets extra training later, usually to improve reasoning and add safety rails. Microsoft’s research team says this one already beats what it calls the strongest open-source base model on the market, DeepSeek-V4-Pro-Base, on internal tests covering science, programming and math.
The bigger claim is efficiency. Microsoft says MAI-Base-2 can be up to 12 times more hardware-efficient by some measures, and that’s not a small tweak. The model uses a mixture-of-experts design, so it only wakes up part of its network for any given prompt instead of lighting up the whole thing. It has 2 trillion parameters in total, but only 80 billion activate on each prompt.
And Microsoft went all-in on that architecture. Its earlier base model mixed MoE layers with FFN modules, but MAI-Base-2 uses MoE layers throughout. The company says that helped the model perform better. It trained the model for 38 days on 16,000 Nvidia GB200 chips, using a dataset of 50 trillion tokens. Roughly a fifth of those tokens went into a computer-vision module, which Microsoft says helps the model read scientific charts better than its predecessor.
The training process got a few careful touches too. Microsoft used gradient clipping to keep training bugs from derailing the run, and Muon, an open-source AI tool, to speed things up. The company plans to license the model’s weights to trusted partners through Azure, then use MAI-Base-2 as the base for a new set of LLMs it says it will talk about soon. So the model is both a research statement and a business move, which is very Microsoft.
My take — AI-written commentary, not fact-checked reporting
This is the kind of model announcement that quietly says more about the industry than the benchmark bragging does. Everyone loves bigger numbers until the bill for the chips arrives, and Microsoft is clearly betting that efficiency is the real prestige feature now. Open source gets to play the benchmark hero; big labs still get to decide who can actually use the weights.
Read more about this at: SiliconANGLE