Leak says Gemini 3 was trained on TPU MoE and is strong on computer use and math

k_kohlbrenner · x · 2026-07-24

A leak posted by @deedydas claims Gemini 3 was shared hours before its official launch and is performing extremely well.

The quoted claims say Google trained it from scratch on TPUs as an MoE model with 1M input and 64k output tokens, which could help keep costs down. The leak also says it looks strong on:

A reply pushes back on the credibility of the poster’s prediction record, but the main item is the leaked model report itself.

Original post →