Microsoft MAI-Base-2: From Reveal to Leaked Specs

Episode 1 · Microsoft AI Unveils MAI-Base-2, a 2.8T-Parameter MoE Foundation Model with Completed Pre-training (2026-10-16, 5 posts)

On October 16, Microsoft AI chief Mustafa Suleyman announced that the MAI team's fully in-house pretraining of a 3-trillion (3T) parameter-scale base model is complete and has reached SOTA level, with post-training results coming soon. The announcement was accompanied by a technical blog post, "Frontier Pre-training at Microsoft AI." This is the first time Microsoft has disclosed training details of a frontier-scale, fully in-house base model, and it is regarded as an important reference for large-model training methodology.

Confirmed

• The model is named MAI-Base-2, built on a MoE architecture with 2.8T total parameters and 80B active parameters; pretraining is fully complete
• Training data scale: 50T tokens
• The official blog discloses extensive pretraining details, with technical commentary from figures such as Elie Bakouch and gordicaleksa
• Elie Bakouch's analysis notes the model is sparser than Kimi K3 (a lower ratio of active parameters)
• Per gordicaleksa's summary of the blog, its FLOPs efficiency is 4-12x higher than DeepSeek
• Accompanying figures show the model leading on multiple internal held-out evaluations; one researcher shared it with the comment "it's a Good Model," noting post-training still needs time

Not yet confirmed

• Evaluation scores and release timing for the post-trained version have not been announced
• The claim of "several times higher pretraining efficiency than DeepSeek" comes from the official blog and awaits independent verification

Why it matters

• This is Microsoft AI's first fully in-house frontier-scale base model; moving away from dependence on external models like OpenAI's is strategically significant
• The disclosed training details (sparsity design, token scale, FLOPs efficiency) offer rare first-hand material on how frontier LLMs are trained — multiple commenters called it a reference for "understanding how large models are trained"
• The 80B/2.8T sparsity ratio shows that MoE sparsification remains the mainstream direction for frontier models

• Microsoft MAI claims SOTA 3T-param in-house base models, post-training results coming — soldni · 2026-10-16
• Microsoft unveils MAI-Base-2: 2.8T-param MoE, 4-12x more pre-training FLOP-efficient than DeepSeek — ChengleiSi · 2026-10-16
• Microsoft unveils in-house 3T-class MAI base models: 80B active params, 50T training tokens detailed — SinclairWang1 · 2026-10-16
• Microsoft's MAI finishes in-house 3T-parameter pretrained base models, post-training results coming — soumitrashukla9 · 2026-10-16
• Microsoft finishes pre-training MAI-Base-2: 2.8T MoE, 4-12x more FLOPs-efficient than DeepSeek — gordic_aleksa · 2026-10-16

Episode 2 · Microsoft MAI-Base-2 Reportedly a 2.8T-Parameter MoE Trained on 16,000 GB200s (2026-10-16, 2 posts)

A leaked spec sheet and Microsoft disclosures indicate MAI-Base-2 is a 2.8T-total/80B-active MoE trained over 38 days on 16,000 GB200 GPUs in Aug-Sep 2026, with an average MFU of just 21%.

• Microsoft's MAI-Base-2: 2.8T-param MoE pretrained on 16k GB200s in just 38 days — scaling01 · 2026-10-16
• Microsoft trained MAI-Base-2 on 16k GB200 GPUs in just 38 days, with 21% MFU — scaling01 · 2026-10-16