Microsoft MAI-Base-2: 2.8 trillion parameters, 80 billion active, not released yet

Microsoft AI says its new MAI-Base-2 is a 2.8 trillion parameter mixture-of-experts model that activates 80 billion parameters per token, trained on 50 trillion tokens in 38 days. It is not released yet, and the headline efficiency claim comes from Microsoft's own tests, so treat it as a credible lab note, not a verified ranking.

The details come from a "Lab Notes" post titled "Frontier Pre-training at Microsoft AI," dated October 15, 2026. It is the clearest look yet at how far Microsoft's in-house model effort has come.

What Microsoft says it built

MAI-Base-2 is a base model, meaning it has finished pre-training but not the post-training that turns it into a chat or agent product. Microsoft describes a mixture-of-experts design with 2.8 trillion total parameters and 80 billion active on any given token. In our arithmetic, that is about 2.9 percent of the weights doing work on each token.

The training mix was 50 trillion tokens, and 18 percent of it was vision-language data, counting repetition. The unique vision-language corpus is 4.5 trillion tokens, about 36 percent more unique tokens than MAI-Base-1. The model is pre-trained natively with vision using what Microsoft calls an encoder-free architecture, so there is no separate image encoder bolted on afterward. Web data has a March 2026 knowledge cutoff.

Microsoft also says every training token was processed in-house, with no distillation from other models. That is a pointed claim in a month when distillation has been the argument of the week, and it is one nobody outside Microsoft can check.

The training run in numbers

The run used 16,000 GB200 GPUs over 38 days in August and September 2026. Average model FLOPs utilization (MFU) was 21 percent and goodput was 89 percent. By our arithmetic, 16,000 GPUs for 38 days is about 14.6 million GPU-hours.

A 21 percent MFU sounds low, and for a dense model it would be. Sparse models spend time routing tokens between experts, which is communication-heavy. Microsoft says it lifted MFU inside the expert layers from 24 percent to 35 percent, which let it use experts in every layer. It also reports that its Muon optimizer added only 2.1 percent, or 165 milliseconds, to the average training iteration.

Goodput is the number worth watching. It is the share of wall-clock time spent making training progress rather than recovering from faults. At 89 percent over 38 days on a new GPU generation, that points to a mature operations team.

The efficiency claim, and what it rests on

Microsoft says MAI-Base-2 is "4-12x more pre-training FLOPs-efficient" than DeepSeek-V4-Pro-Base, with the range given later as 4.2 to 12.3 times. Against its own MAI-Base-1, it reports a 6.3 to 12.7 times improvement. The comparison set also includes Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, and Microsoft notes that some of those were post-trained checkpoints, not clean base models.

DeepSeek's Hugging Face listing for V4-Pro-Base shows 1.6 trillion parameters, so MAI-Base-2 is much larger in total size. The note does not publish specific benchmark scores in its text. They sit in figures that measure held-out text loss (bits per byte), code, STEM, math and question answering.

Here is the hype check. A FLOPs-efficiency ratio depends on which losses you match, which tokens you count and how you charge for a larger model's extra capacity. A range of 4.2 to 12.3 times is wide enough that the true answer depends heavily on those choices. And "outperforms every open-weight model we tested" is bounded by the words "we tested."

Why a base model is a different thing from a product

Nothing here ships today. Microsoft says MAI-Base-2 "is currently undergoing post-training, and we look forward to sharing more results soon." There is no release date, no price and no API.

Microsoft's stated plan is to license the weights to trusted partners through Foundry. That is neither an open release nor a public API. It sits between the two, closer to how Microsoft already distributes models to enterprise customers, and it keeps the weights out of the open-weight comparison it just won.

For readers tracking the race, the interesting fact is the cost structure. Microsoft is claiming it can pre-train a frontier-scale sparse model on a single 16,000-GPU cluster in about five and a half weeks. If the efficiency claim holds up even at the low end of the range, rivals on older recipes pay more for the same capability. Our earlier look at Microsoft's Maia 300 rack covers the hardware side of that bet.

What to watch

Three things will tell us whether this is a frontier model or a strong base. First, the post-trained version: base-model loss numbers rarely predict agent performance. Second, an independent evaluator such as Artificial Analysis or Epoch AI running it on shared tasks. Third, the Foundry terms, since license conditions decide who can actually use it.

How it compares with what we already know

Microsoft's note arrives in the same week as Meta's Muse Spark Open and Nvidia's Nemotron 3.5 line, and all three are sparse designs. Meta's model has 816 billion total and 42 billion active parameters, so MAI-Base-2 is roughly three and a half times larger in total size and about twice as large in active size. Size alone proves little, which is why the held-out evaluations matter more than the parameter count. The useful question for the next few weeks is whether Microsoft publishes the evaluation figures behind its charts, along with the unique-token counts for each data source, so outsiders can reproduce even a slice of the comparison.

Our take

This is a strong engineering disclosure and a weak competitive proof. The run statistics are specific and plausible, and the no-distillation claim, if true, matters. But every comparison was chosen and run by Microsoft, the model cannot be downloaded, and its post-training has not been shown. We would treat the 4 to 12 times figure as Microsoft's estimate and wait for a post-trained model with outside scores before moving MAI above the open-weight leaders in any ranking.