Microsoft unveils a 2.8-trillion-parameter AI that nobody can use yet

Sixteen thousand NVIDIA GB200 GPUs, thirty-eight days of computing in August and September 2026. That is what it took Microsoft AI, Microsoft's AI division, to pre-train MAI-Base-2, presented on October 15 in a technical blog post.

The model has 2.8 trillion parameters in total, of which 80 billion are active for each word processed. It relies on a "mixture of experts" architecture: only part of the network works on each piece of text, which reduces the computing cost.

More efficient than DeepSeek? According to Microsoft

The company claims its training method would be 4.2 to 12.3 times more compute-efficient than that of DeepSeek-V4-Pro-Base, which it considers the best-performing open-weight base model. The gain would reach 6.3 to 12.7 times compared with its predecessor, MAI-Base-1.

These figures are estimates, not direct measurements. On four unpublished test sets, Microsoft plotted the curve linking computing power to the model's prediction quality, then deduced the compute needed to match DeepSeek. The coding tests come from its own source code. The science, math and question-answering tests were commissioned from contractors.

The post acknowledges a limitation: against Kimi K3, Qwen3.8 or MiMo-V2.6, Microsoft could only compare with versions that had already been fine-tuned, which may skew the results. The specialist site byteiota considers the comparison unfair. No independent evaluation is available to date.

What does the model contain?

MAI-Base-2 was trained on a mix of 50 trillion text elements (tokens), 18% of which are data combining text and images, with web data cut off in March 2026. Microsoft says its pre-training relies on "organic" data, without distillation from other models. Its web crawler reportedly respects the robots.txt file, through which websites refuse data collection. The post does not address the question of copyright.

Why you won't be able to try it

MAI-Base-2 is a base model. It still has to go through the fine-tuning stage that will turn it into a usable assistant. Microsoft has announced no date, no price and no public access. It plans to license the weights to "trusted partners" through Foundry, its platform for businesses. So this is not an open model.

The contrast is stark with NVIDIA or Meta, which publish their weights. For a European CIO, this largely strategic announcement says one thing: Microsoft wants its own frontier models and control over their entire manufacturing chain. Microsoft says it is already preparing MAI-Base-3. The real tests will come with the fine-tuned version, if independent evaluators get access to it.