Microsoft Previews MAI-Base-2, a 2.8-Trillion-Parameter Model It Will Not Open Up
Microsoft's MAI-Base-2 is a 2.8T-parameter foundation model trained on 50T tokens that Microsoft says learns far more efficiently than DeepSeek's. It is still in training and will go only to trusted partners.
Microsoft AI has described MAI-Base-2, a very large “base model” with 2.8 trillion parameters in total and 80 billion active at a time. A base model is the raw first stage of an AI: it has read an enormous amount of text and learned patterns, but has not yet been taught to chat politely or follow instructions. That second stage, called post-training, is still underway.
The details
Microsoft says MAI-Base-2 was trained on 50 trillion tokens, small chunks of text that serve as the unit of AI reading. Its pool of unique data grew about 36 percent over the previous version, and 18 percent of the mix was images. The headline claim is efficiency: Microsoft says it is 4 to 12 times more efficient in pre-training compute than DeepSeek-V4-Pro-Base, and that it beats every open-weight model the team tested on its own held-out exams.
The fine print matters. The efficiency numbers come from fitted scaling formulas, not from training two models side by side. Several rival models were only available in already post-trained form, which can skew comparisons. Code tests used Microsoft’s internal codebase. And final performance may change before release.
What is behind this
For years Microsoft leaned on OpenAI’s models. Building its own foundation model is a bid for independence, and efficiency is the sales pitch: if you need a fraction of the computing power to reach the same quality, your costs and energy bills drop. Microsoft plans to license the weights to trusted partners through its Foundry platform, with no date, so unlike Meta’s and Nvidia’s releases this week it will not be openly downloadable.
What this means for you: Nothing you can use today. It is a signal worth tracking for how the AI race is shifting from “who has the biggest model” to “who trains smartest.” Wait for post-training results and independent tests before drawing conclusions.
Sources
Source: https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/
Meta's Muse Spark Open Is a Huge Open-Weight Model Built for Two Mac Studios
Meta released Muse Spark Open, an 816-billion-parameter open-weight model with a one-million-token window, and says it runs on two Apple M5 Ultra Mac Studios. Benchmarks and the full safety report are still missing.
