Skip to content
Neural Daily

ResearchNews

MAI-Base-2: Microsoft finishes 2.8T-parameter base model

Microsoft AI said on October 15 that it has finished pre-training MAI-Base-2. The 2.8-trillion-parameter MoE base model activates 80 billion parameters per token and, according to Microsoft's own held-out tests, beats every open-weight model it evaluated.

By: Mikola Máté Patrik · · 2 min read

Abstract header image from Microsoft AI frontier pre-training report
Image: Microsoft AI

What Microsoft announced

MAI-Base-2 succeeds MAI-Base-1: 2.8 trillion total parameters, 80 billion active. It was trained on 16,000 GB200 GPUs for 38 days in August and September 2026, on about 50 trillion tokens (Microsoft AI). It is now in post-training and not yet available as a product.

Four bar charts of held-out bits per byte on code, QA, STEM and math, where MAI-Base-2 scores lowest
Microsoft AI

Source: Microsoft AI. Held-out bits per byte on code, QA, STEM and math evaluations; lower is better. MiMo, Qwen and Kimi results are from post-trained checkpoints.

Efficiency claims

Microsoft says its recipe is 4.2–12.3x more FLOPs-efficient than DeepSeek-V4-Pro-Base, the strongest open base model, and 6.3–12.7x better than MAI-Base-1. These are Microsoft's private held-out evaluations, not public benchmarks.

Business Insider reported that Mustafa Suleyman told staff the model beats a comparable OpenAI base model with about a quarter of the compute. That comparison is also internal.

Methods in brief

  • Scaling ladders with 'source scaling' so small-scale results transfer to large runs.
  • MoE in every layer with a custom dropless implementation; MoE-layer MFU rose from 24% to 35%.
  • Muon instead of AdamW, plus activation gradient clipping for stability.
  • Native, encoder-free vision; visual data made up 18% of the 50T-token mix.

Average MFU over the run was 21% and goodput 89%. Microsoft stresses that its data is organic rather than distilled from other models, which it says lets it license weights to partners through Foundry.

Why it matters

It is the clearest signal yet that Microsoft is building its own frontier model alongside its OpenAI partnership. What that means in practice will show once the post-trained model faces public evaluations. Microsoft says MAI-Base-3 preparation is already under way.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles