# Distribution modifications

Prepared by FreeVideo from OpenVDN/vdn-minimax-h3@751739ee5b9e3ac802dca5d5111075fdaeb47885. Powered by MiniMax H3.

The FP8 block and root archives are byte-for-byte identical to the earlier full Edge package and retain their embedded modification notices. Eligible linear weights use per-tensor FP8 scales; other parameters retain their source precision. Unused BF16 backup weights are omitted.

The slim format removes original AdaLN projection groups from the default payload and makes their fixed eight-step modulation outputs required model assets. These table archives retain their original tensor bytes and embedded notices. Their directory/catalog identity now binds the weights and exact schedule, while the original GPU and software identity is retained separately as provenance. Identical visual-keyframe schedules share one table set.

The optional original projection source is pinned to the prior complete repository revision. It is not part of the default download. A modulation-changing adapter or unsupported schedule may request it explicitly through the engine's compatibility path.

The new catalog, identity records, manifest and model card describe this format change. LICENSE is an unmodified copy of the upstream agreement; NOTICE preserves the upstream attribution.

## Rowwise variant

`rowwise/cache/` adds the existing official per-channel FP8 weight format. The 51 block/root archives carry modification notices in their safetensors metadata. Their tensor payloads are copied unchanged from the previously prepared rowwise cache; no requantization or training occurs during packaging. The portable modulation constants match the same original projection and time-embedding weights. Existing `cache/` per-tensor files and original licenses are unchanged.