2026-10-11 17:09 UTC

distillation

band: coolmomentum: stable score: 0.012
temperature history

Episodes (2)

Independent evaluations will determine whether preserving internal representation geometry during quantization-aware distillation improves NVFP4 model quality over KL-only distillation without reducing low-precision efficiency.
expiredconvergesscott: medium
Video Rebirth claims its released HyperFlow LoRA reproduces MiniMax-H3's 49-forward sampling trajectory in eight steps using only the base model as teacher, potentially lowering video-generation inference costs without external training data.
seed