MAI-Base-2: Microsoft's 2.8T Parameter Pre-Training Run

A finished pre-training run, not a release

On Thursday, October 15, 2026, Microsoft AI published a technical post saying it had completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active. It was trained on 16,000 NVIDIA GB200 GPUs over 38 days in August and September 2026, at an average model FLOPs utilization of 21 percent and 89 percent goodput. Its predecessor, MAI-Base-1, underlies the MAI-Thinking-1 model Microsoft announced in June. [1][2]

Nothing was released. MAI-Base-2 was still in post-training, with results promised later, and Microsoft said preparations for MAI-Base-3 had begun. SiliconANGLE's report of the post, published that evening US time, gave the size as 2 trillion parameters; Microsoft's post says 2.8 trillion. [1][2]

How Microsoft compared it

Microsoft measured bits per byte on held-out text, which does not depend on each model's tokenizer: code from its own codebase, and STEM, math and question-answering sets it commissioned from vendors and decontaminated against its training data. On those sets MAI-Base-2 scored better than every open-weight model it tested: DeepSeek-V4-Pro-Base, which Microsoft calls the strongest open base model, plus Kimi K3, Qwen3.8 2.4T and MiMo-V2.6-Pro. Only post-trained checkpoints of the last three exist, and Microsoft noted that post-training may affect the comparison. [1]

Its efficiency claim is an estimate. Fitting scaling laws to its own models, Microsoft calculated how much compute its recipe would need to match DeepSeek V4 Pro Base and concluded the recipe is 4.2 to 12.3 times more efficient in training FLOPs, and 6.3 to 12.7 times more efficient than its own recipe for MAI-Base-1 in June. It credits data quality and modeling changes, including mixture-of-experts blocks in every layer rather than every other one, faster MoE and Muon optimizer implementations and scaling ladders that account for repeated data. [1]

The comparison is narrow by design. A base model is scored on how well it predicts held-out text, which does not show how the finished model will do on coding, agent or reasoning tests, and no closed frontier model appears in the published table. What the post does establish is that Microsoft ran a frontier-scale pre-training run on its own. [1]

The route to customers

Microsoft said its pre-training uses organic data with no distillation from other models, and that this provenance will let it license the model weights to trusted partners through Foundry. It also said it would tune models and infrastructure together, first for its own products and then for Foundry customers, which connects to the Frontier Tuning service it previewed in June for training models inside a customer's own environment. No licensing terms, prices or dates were given. [1][2]