Announcement
DeepSeek-V4-Pro-Base’s pre-trained score reportedly matched on held-out code with 12x less training compute
The poster says MAI-Base-2 trained cleanly throughout its run, crediting careful scaling ladders and recipe refinement.
TLDR
A poster claims they matched DeepSeek-V4-Pro-Base’s pre-trained score on held-out code using 12x less training compute. In a reply, the same poster says MAI-Base-2 trained cleanly over its full run, crediting careful scaling ladders and steady refinement of the recipe.
Combined views
16.8K
2 Sources, first seen ago
