mlx-community/gemma-4-E4B-it-qat-assistant-bf16
Same weights, converted vs google/gemma-4-E4B-it-qat-q4_0-unquantized-assistant. Same weights stored in another dtype (BF16/I64 → BF16/I32).
Tensors changed: Embeddings 0/3, Norms 0/21, other 0/6, MLP 0/12, Attention 0/8
1 other repos hold byte-for-byte identical weight files.