Microsoft AI finishes pre-training MAI-Base-2: a 2.8T-parameter MoE (80B active) it says beats every open base model it tested

Why it matters

【Key points】

In a technical post on October 15, Microsoft AI (MAI) said it has finished pre-training MAI-Base-2, an MoE base model with 2.8T total and 80B active parameters. It trained on 16k GB200 GPUs for 38 days in August and September 2026, at 21% average MFU and 89% goodput.

On held-out evaluations (MAI's own codebase plus vendor-commissioned STEM, math and Q&A sets, scored in bits-per-byte), MAI-Base-2 scores better than DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL; Microsoft notes the last three are only available as post-trained checkpoints. It estimates its recipe is 4.2–12.3x more FLOPs-efficient than DeepSeek-V4-Pro-Base and 6.3–12.7x better than June's MAI-Base-1. The post says pre-training used organic data with no distillation from other models, and Microsoft plans to license the weights to trusted partners through Foundry. Post-training is under way; Mustafa Suleyman says results are coming soon.

【Verifiable】

Microsoft AI technical post: https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/

Mustafa Suleyman's post: https://x.com/mustafasuleyman/status/2110766651344146564

【Our read】

1. This is Microsoft's clearest step toward an in-house frontier model: 2.8T parameters, a 16k-GPU run and a fully owned stack, a serious hedge beyond OpenAI.

2. So far there is only a base model and bits-per-byte, which measures text prediction, not task performance; the real comparison needs the post-trained model and third-party evals.

3. Licensing weights to trusted partners is not open weights; watch whether Foundry customers can self-host and at what price.

4. The emphasis on no distillation and clean data provenance is a compliance pitch to enterprises that may matter commercially more than the scores.