Oh boy! Its another copy of schnell with the double blocks from dev! So what's the difference? I didn't change anything besides the blocks required to make the model fast. It's intended to be trained on top of, quantized and merged. No other layers were touched. No FP8's, No NF4s. Nothing else included. Just model.Tested at full precision and the text capabilities were mostly intact.

Flux Pure Merge BF16

Quant it to your favorite precision. Train on it, etc.

Files with SHA256: 62b66b45074309d3a35b7941c9d8745334cb708e4c6e6b687e82ab36b52187b3