xiaomi-research/MiLMMT-46-4B-Pretrain

Full fine-tune vs google/gemma-3-4b-pt. Fine-tune: 90% of the weights changed, across 34 of 34 layers.

Tensors changed: Embeddings 1/4, Norms 204/316, MLP 102/210, Attention 136/352, other 0/1

1 derivatives checked