i've got an idea!
#23
by soupcheese - opened
lets bloat the model with a giant text encoder and make it untrainable to most people. these are supposedly the models for the poors so why not just use the same text encoder as ZiT and the 4B model and actually let people play with it? done with black forest, I hope all the ZiT's and ZiB's and every chinese model overshadows you.
soupcheese changed discussion status to closed
soupcheese changed discussion status to open
i've seen people using the clip "qwen_3_8b_fp8mixed.safetensors" and select type "flux",
edicamargo/qwen_3_8b_fp8mixed_abliterated, 7.62 gb, and works. (just verified and working)
