MiniMax-H3 bX-49 β€” full hybrid teachers

Full (non-pruned) MiniMax-H3 checkpoints built the same way as the popular b25-49 pruned hybrids: FL2VA is the base for the whole network, and the adaln_proj of blocks X-49 is taken from Ref2VA.

Why: a pruned bX-49 file cannot be recorded from β€” it has no time_embedder and its AdaLN is an 8-wide curve β€” so the exact modulation has to be reconstructed. These files carry the whole AdaLN branch at full width. One recording pass on one of them yields that hybrid's native modulation, and gives a bitwise end-to-end reference: pruned bX-49 + table recorded here reproduces this file exactly.

Files

file split replaced tensors size sha256
minimax_h3_hybrid_fl2va_ref2va_b5-49_full_int8.safetensors 5 180 31.7 GiB 5100ee083264653c8f49dd5fdccd8a591a454810b565f64ac2e323c9e80f79a5
minimax_h3_hybrid_fl2va_ref2va_b20-49_full_int8.safetensors 20 120 31.7 GiB bae89706229fda2d6523a9ad028b2729f53b36be0cf224fe641c9a4993e84427
minimax_h3_hybrid_fl2va_ref2va_b25-49_full_int8.safetensors 25 100 31.7 GiB 35a57db840bebe701b151cb9ea8c563db5b2d267927b283ad54162ee6418f84f
minimax_h3_hybrid_fl2va_ref2va_b30-49_full_int8.safetensors 30 80 31.7 GiB 7f3e72d1479ded3b9880a10d838e22a3791e1c3be9a137fd5cb19a3f7f3550d5
minimax_h3_hybrid_fl2va_ref2va_b45-49_full_int8.safetensors 45 20 31.7 GiB e6538c349820db0d97c4e74c5ae8ba4f970e5a32fe100f6a2bcfa28a7562f5f6

Every file is a byte copy of fl2va_int8_convrot with the AdaLN block range swapped in from ref2va_int8_convrot (4 tensors per block: weight I8, weight_scale F32, bias BF16, comfy_quant U8). Tensor shapes and all data_offsets entries are unchanged, so the FL2VA header stays valid.

Sources: Comfy-Org/MiniMax-H3.

Verification (byte level, every file)

  • every replaced tensor is byte-identical to the Ref2VA source
  • every other byte of the body is byte-identical to the FL2VA source
  • sha256 above

End-to-end test

b25-49, both chains with the same LoRA, 8-step simple, shift 12/2, 832x480x39:

A: full b25-49 hybrid (native)
B: pruned b25-49 + table recorded on A
samples__part0: equal=True max_abs=0.000e+00 psnr=inf
samples__part1: equal=True max_abs=0.000e+00 psnr=inf
VERDICT: PASS - bitwise identical

Which split should I use?

Pick the one that matches the pruned file you already run. A pruned bX-49 keeps the base teacher's body and only adopts the donor's AdaLN from block X on, so the split is a property of that file, not a quality dial:

  • low X (5, 20) β€” most of the network modulated by Ref2VA: strongest reference conditioning, largest deviation from pure FL behaviour;
  • 25 β€” the common b25-49 recipe;
  • high X (30, 45) β€” only the tail adopts Ref2VA, closest to plain FL.

This is NOT the same target as the two-teacher merge

Recording FL2VA and Ref2VA separately and merging the tables gives a different table for the tail blocks: FL and Ref do not share a time_embedder (proj_out.weight differs by up to 0.104), so the merged tail is Ref's modulation computed with Ref's own embedding, while the tail blocks of a real hybrid see the base teacher's embedding. Measured on b25-49: blocks 0-24 and final_layer bitwise identical, blocks 25-49 differ by up to 0.19; applying the merged table to the pruned file and comparing against this model gives psnr=22.20 dB / 31.10 dB.

To reach the same target without a merged checkpoint, record Ref2VA full with H3AdaLNRecord.time_embedder_from = <FL2VA checkpoint> and merge that table β€” it matches these files bitwise (25/25 tail projections, 51/51 after the merge).

Usage

  1. ComfyUI/models/diffusion_models/; the stock MiniMax-H3 loader accepts them (int8 convrot, same as the official full builds).
  2. With ComfyUI-H3-ExactAdaLN: load one (+ your LoRA) -> H3 AdaLN Record before the guider -> run the schedule you will use -> H3 Exact AdaLN Save Table -> on the matching pruned bX-49, H3 Exact AdaLN Apply. One teacher, one recording, no merge.

A table recorded with one LoRA reproduces the model with a different non-AdaLN LoRA bitwise; only LoRAs that touch time_embedder / adaln_proj (acceleration / distillation LoRAs) need a re-record.

Limits

  • A table is a function of (weights, LoRA stack and strengths, sigma grid, shifts, packed layout); change any of them and it must be re-recorded.
  • out_dtype is fixed at recording time; these models run bf16 maths.
  • Derivative of MiniMax H3: the merge and quantisation are Apache 2.0, the underlying weights remain under MiniMax's own license.
Downloads last month
628
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for binglingzhimeng/MiniMax-H3-b25-49-full

Finetuned
(30)
this model