# Microsoft AI pre-trained MAI-Base-2 with 2.8 trillion total parameters > Source: > Published: 2026-10-17 08:02:04.568414+00:00 [Hacker News](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/) ### Microsoft AI pre-trained MAI-Base-2 with 2.8 trillion total parameters Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Microsoft now has a fully in-house frontier base model: MAI-Base-2, an 80B-active/2.8T-total MoE trained from scratch on organic (non-distilled) data, claiming to beat every open-weight model tested and run 4-12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The key practical angle is licensing: because of clean data provenance, Microsoft plans to offer these weights to enterprises via Foundry, giving you a legally-defensible, self-hostable alternative to Llama/DeepSeek rather than just another API — though it's still in post-training, so the Thinking/Code/Voice derivatives are what you'll actually ship against. Microsoft has completed pre-training MAI-Base-2, an MoE foundation model with 80B active and 2.8T total parameters, trained on 16k GB200 GPUs over 38 days. Microsoft claims it beats every open-weight model it tested and is 4–12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, positioning it as a proprietary, clean-provenance base model for post-training and eventual licensing through Foundry. The practical consequence is that Microsoft is trying to own more of the frontier model stack instead of depending on partners, with enterprise weight licensing as the likely product angle once post-training is done. ### AI vs. AI Debate “The summary is mostly accurate but overstates the unsupported “self-hostable alternative” claim and omits the notable training-scale and systems metrics: 16k GB200s, 38 days, 21% MFU, and 89% goodput.” “The "self-hostable" framing follows directly from Microsoft's stated plan to offer enterprises the actual weights via Foundry, which is precisely what distinguishes this from an API and is the whole licensing point I emphasized, while the training-scale metrics Model B lists are secondary infrastructure details that don't change the article's practical takeaway.”