Microsoft Finishes Pre-Training MAI-Base-2, a 2.8 Trillion Parameter Model It Plans to License

Microsoft AI, the company's in-house model lab, said on October 15, 2026 that it has finished pre-training MAI-Base-2, a mixture of experts model with 80 billion active and 2.8 trillion total parameters. The model is still in post-training and is not available yet; Microsoft says it plans to license the weights to trusted partners through its Foundry platform.

The report follows MAI-Thinking-1, the reasoning model Microsoft introduced at its Build conference in June 2026 (Pure AI), which was built on the smaller MAI-Base-1 with 35 billion active and about 1 trillion total parameters. Microsoft opened a public preview of MAI-Thinking-1 in Foundry, according to its announcement page, updated on August 12, 2026. MAI-Base-2 takes Microsoft's model releases into the size class of the largest open weight models from Chinese labs, which is where it aims its comparisons.

What did Microsoft AI announce?

Microsoft AI published a technical report, not a product: MAI-Base-2 is a base model, the raw result of pre-training, before the post-training that teaches a model to follow instructions and reason. Microsoft says the base model will be the foundation of its next generation of models and that preparations for MAI-Base-3 are already underway.

Pre-training ran on 16,000 Nvidia GB200 GPUs for 38 days in August and September 2026, according to the report. Microsoft reports an average model FLOPs utilization of 21%, a measure of how much of the chips' theoretical compute went into useful training math, and says training progressed during 89% of the run.

The training mixture held 50 trillion tokens, up from 30 trillion for MAI-Base-1, with a knowledge cutoff of March 2026. Microsoft says its pool of unique training tokens grew by about 36%. Unlike its predecessor, MAI-Base-2 learned from images from the start: vision data made up 18% of the mixture, drawn from 4.5 trillion unique tokens of image and text material.

The architecture also changed. MAI-Base-2 places expert layers in every one of its 102 layers instead of alternating them with dense layers, and trains most of its weights with the Muon optimizer instead of AdamW, which Microsoft says gave more stable training and large efficiency gains.

How does MAI-Base-2 compare with open weight models?

Microsoft says MAI-Base-2 beats every open weight model it tested on its held-out evaluations, and its chart shows the lowest score of six models on all four tests. The tests measure bits per byte, how well a model predicts unseen text regardless of its tokenizer, where lower is better.

The comparison has limits that Microsoft spells out. Only DeepSeek-V4-Pro-Base is a released base model; Kimi K3, Qwen3.8 and MiMo are available only as post-trained checkpoints, and Microsoft notes that post-training may affect these scores. The code test uses Microsoft AI's own codebase, while the STEM, math and QA sets were commissioned from vendors to limit contamination, so none of the four can be checked by outsiders.

Kimi K3, which Moonshot AI lists with the same 2.8 trillion total parameters on its model card, comes second on all four tests and trails MAI-Base-2 by 0.0028 bits per byte on math. The chart includes no closed models from OpenAI, Anthropic or Google.

What does the 4 to 12 times efficiency claim mean?

Microsoft estimates that its training recipe would need 4.2 to 12.3 times less compute than DeepSeek used to reach DeepSeek-V4-Pro-Base's scores, depending on the test. The figures come from scaling laws fitted to Microsoft's own smaller models, not from a run that matched DeepSeek's model.

By test, the estimate is 12.3 times for code, 4.2 for QA, 8.3 for STEM and 6.6 for math. Against its own MAI-Base-1, released in June, Microsoft reports a gain of 6.3 to 12.7 times.

MAI-Base-2 itself did not use less compute than DeepSeek's model. DeepSeek says on its model card that it pre-trained V4-Pro, with 49 billion active parameters, on more than 32 trillion tokens; MAI-Base-2 has 80 billion active parameters and trained on 50 trillion. Microsoft says DeepSeek's model was trained on "significantly fewer FLOPs" and that it spent the efficiency gain on a larger model.

Who can use MAI-Base-2, and when?

No one outside Microsoft can use MAI-Base-2 as of October 2026, and Microsoft gives no release date. It says it will tune its models and infrastructure first for its own products and then for Foundry customers.

The run used Nvidia hardware, while Microsoft said on October 12, 2026 that its own Maia chips will become the primary inference system for its products from 2027. Microsoft has also said it will rebase its Decision-1 classifier on MAI models, among others.

What we don't know yet

• The name, release date and benchmark results of the post-trained model built on MAI-Base-2.
• Which companies qualify as trusted partners for the weights, and on what license terms and prices.
• How MAI-Base-2 scores on public benchmarks or in independent tests, and against closed models.
• Whether Microsoft will publish its held-out evaluation sets so others can reproduce the comparison.