MiniMax H3 FL vs REF checkpoints and the Cobijada weight-selection hybrid

MiniMax H3 is a joint audio and video diffusion transformer released as two checkpoints with the same architecture and tensor layout: FL, conditioned on first and last keyframes, with higher visual and audio quality but no reference-driven generation; and REF, which adds image, video and audio reference conditioning at a cost to output quality, including on non-reference tasks. The community Cobijada hybrid (Winnougan) reports that most weights are effectively identical between the two and makes a tensor-level weight selection: the FL checkpoint everywhere, except the adaln_proj weights of transformer blocks 25 to 49, taken from REF to keep its reference pathway. No training or fine-tuning was performed. Both sources were pruned variants that the card labels uncensored. Files: BF16 (full-precision merge), INT8 ConvRot and W4A8 ConvRot quantizations, all pruned. In ComfyUI, place the file in models/diffusion_models and load it with the normal MiniMax H3 checkpoint loader; no custom node is needed. Both reference-conditioned and first/last-keyframe workflows use the same checkpoint. The author does not expect it to exceed FL on non-reference generation. Rights: the repository declares Apache-2.0 for the merge and conversions, but the card states that the underlying MiniMax H3 weights remain subject to MiniMax's own licence; review the MiniMax H3 terms before use. Based on the Winnougan/Cobijada_Minimax-H3_Hybrid_Pruned_ComfyUI card pinned to revision d14d58723efadf84de98bc2c3a3613cfedea4d04 and observed on 2026-10-27. Nothing was run here; the weight-similarity analysis and quality claims are the merger's, and the MiniMax H3 licence was not reviewed. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/Winnougan/Cobijada_Minimax-H3_Hybrid_Pruned_ComfyUI/raw/d14d58723efadf84de98bc2c3a3613cfedea4d04/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 632:725, 1596:1975, 2917:2935, 3839:3849; exact excerpt SHA256 ce0ea2c034c0cd04aab722ca58f90b3816b4c67be8e94c0b2f9a9c0701062eb3,4cfac40924a89a0915ea4df16145bc517b63f93ea992ed762f76cc2f36e0f422,bd605917c315dad90d300041b642f0899e779d062220ba0a6d2be60ce30a3fcc,6c549a56db01e021f248"
    }
  ]
}