Frontier Pre-Training at Microsoft AI
What happened
Microsoft announced the completion of pre‑training its new foundation model, MAI‑Base‑2, an 80‑billion‑parameter mixture‑of‑experts (MoE) model with 2.8 trillion total parameters. The model is 4‑12 times more FLOPs‑efficient in pre‑training than the strongest open‑weight competitor, DeepSeek‑V4‑Pro‑Base, and outperforms every open‑weight model tested on Microsoft’s held‑out evaluations. Training used 16,000 GB200 GPUs over 38 days, achieving an average model FLOPs utilization of 21 % and a goodput of 89 %. Microsoft plans to license the model weights through its Foundry service after post‑training is finished.
Why it matters
The efficiency gains could lower the cost of delivering advanced AI capabilities to enterprises, making powerful language models more accessible. Users may see improved AI features in Microsoft products that rely on these models.
Reality check
An automated, sceptical read of the sourceNotable. The model shows significant efficiency and performance improvements over existing open‑weight models, but it builds on known scaling techniques rather than introducing a wholly new breakthrough.
Is this a big deal?
Get stories like this every morning
The day's AI news in plain English, in your inbox at 8:30 IST. Free.
More from Microsoft
Microsoft Workers Call for Accountable AIMicrosoft · 13 Nov 2026The end of Microsoft Research's AI Frontiers groupMicrosoft · 7 Nov 2026Microsoft is winding down Oracle deal to rent AI cloud capacityMicrosoft · 6 Nov 2026Microsoft and Epic Games Suddenly Agree That AI Is TheftMicrosoft · 28 Oct 2026New AI Is More Than Twice as Good at Diagnosing Diseases as DoctorsMicrosoft · 23 Oct 2026This page was collected from microsoft.ai and summarised automatically by AIPulsee's pipeline. See how it runs.