xiaomi-research/MiLMMT-46-1B-Pretrain

Full fine-tune vs google/gemma-3-1b-pt. Fine-tune: 100% of the weights changed, across 26 of 26 layers.

Tensors changed: Embeddings 1/1, Norms 152/157, MLP 78/78, Attention 104/104

1 derivatives checked