Mustafa Suleyman: MAI now has state-of-the-art pre-trained base models ‘in the 3T-parameter weight class’
Mustafa Suleyman @mustafasuleymanX
Why it matters
Microsoft AI CEO’s announcement of MAI-Base-2; ~1.2M views, the highest-reach post in the Oct 16 sweep’s X list.
Summary
Announcement post (~1.2M views, ~3.1k likes by Oct 16 per fxtwitter): ‘We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon -- stay tuned!’ Links the technical post microsoft.ai/news/frontier-pre-training-at-microsoft-ai/. A follow-up (https://x.com/mustafasuleyman/status/2110766654863167694, ~16k views) says: ‘On held-out code, we matched DeepSeek-V4-Pro-Base’s pre-trained score using 12x less training compute.’
Archived text
We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI.
Post-training results coming soon -- stay tuned!
Lots of details in our technical blog post.
https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/