Microsoft AI Says MAI-Base-2 Has 2.8T Parameters and Trained on 16,000 GB200s in 38 Days

Microsoft AI said on Thursday that it trained MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token, on 16,000 Nvidia GB200 GPUs over 38 days. The company has not released the model.

The figures come from a Microsoft AI "Lab Notes" post titled Frontier Pre-training at Microsoft AI, dated October 15. The post names no author. It says MAI-Base-2 is still in post-training and gives no release date.

What Microsoft says it built

MAI-Base-2 is larger and sparser than MAI-Base-1, the base model behind MAI-Thinking-1, which Microsoft announced in June. About 2.9 percent of its parameters are active on any token. The post says the model is trained natively on images with an encoder-free vision architecture.

The training mixture was 50 trillion tokens. Microsoft said the pool of unique tokens grew about 36 percent from MAI-Base-1, and that its vision corpus holds 4.5 trillion unique tokens, or 18 percent of the mixture once repeats are counted.

Item	Figure in the post
Total / active parameters	2.8T / 80B
Training tokens	50T
Hardware and duration	16k GB200, 38 days
Average MFU / goodput	21% / 89%

MFU is the share of the chip's theoretical peak that training actually used. Microsoft normalises it to 2.5 petaflops per GB200, counts neither rematerialisation nor optimizer work, and reports 21 percent on average. It said mixture-of-experts layers rose from 24 percent to 35 percent MFU, which let it put experts in every layer. MAI-Base-1 used them in every other layer.

The efficiency claim is Microsoft's own

The headline comparison is against DeepSeek-V4-Pro-Base. Microsoft says MAI-Base-2 is 4.2 to 12.3 times more pre-training FLOPs-efficient, and 6.3 to 12.7 times more efficient than MAI-Base-1. Both are vendor-supplied figures. The post does not say how the range was derived.

DeepSeek's Hugging Face page lists 1.6 trillion parameters for that model and carries no model card. It gives no active-parameter count, token count or licence, so the denominator in Microsoft's ratio cannot be checked from public material.

1. MAI-Base-22800 B
2. DeepSeek-V4-Pro-Base1600 B

Source: Microsoft AI Lab Notes, Oct. 15, 2026; Hugging Face model page for DeepSeek-V4-Pro-Base, accessed 2026-10-16

Microsoft also compared against Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. It evaluated those three as post-trained checkpoints and said post-training may affect the results. Its post says MAI-Base-2 "outperforms every open-weight model we tested on our held-out evaluations."

The held-out sets are not public. Microsoft said code evaluations use its own codebase, and that STEM, math and question-answering sets were commissioned from vendors. All were decontaminated against the training data, according to the post. No outside party is named as having run any of them.

Data and licensing

Microsoft said the data is organic, with no distillation from other models, and that every training token was processed in-house. Its web data has a March 2026 cutoff, and its crawler respects robots.txt. It excludes sources that break Microsoft's Responsible AI policies or appear on the USTR Notorious Markets list.

The weights will not be public. The post says Microsoft plans to license MAI-Base-2 to trusted partners through Foundry, and describes that as a future capability. That is a different position from Meta's Muse Spark Open, whose 816 billion parameter weights are downloadable, and from the small open releases such as Liquid AI's LFM2.5-24B-A2B.

What is still unknown

Microsoft did not publish benchmark scores for MAI-Base-2, only that it ranks ahead of the models it tested. It did not state the cost of the run, the licence terms for partners, or when post-training ends.

Microsoft said preparations for MAI-Base-3 are under way and that it will share more MAI-Base-2 results, with no date given.