See-Krea2

FP8 E4M3, per-tensor scaled — 13.1 GB. The build with the widest loader support, and the format Krea themselves ship.Output-side projections (attention output, attention gate, MLP down) are flagged for full-precision matrix multiply, so error does not accumulate where it matters most; the input projections use the fast FP8 path. That split is not my invention — it is the policy read directly out of Krea's own official FP8 release.Measured weight deviation: 2.66 %. That is simply what FP8 E4M3 costs — three mantissa bits do not give more, and Krea's own FP8 build measures the same.Runs on: RTX 40xx (Ada) and newer for the native FP8 path. Older cards can still load it with software dequantisation. Pick this if: you are on Ada, or if another build fails to load and you want the safest option.Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP.

See-Krea2-Genesis_fp8_scaled.safetensors

Genesis-fp8_scaled

Genesis-V2

Text Encoders

Genesis-int4_convrot

Genesis-nvfp4

Genesis-int8

Genesis-int8_convrot

Genesis-BF16

Files with SHA256: c90f03ab41bd135c4a24bc7c2c359b04d1f8a79a7c8b2f97764e8e35fac4ee88