MiniMax-H3 Pruned Ref-Delta Fused r1024 — ComfyUI Single File
Native ComfyUI-format single-file conversion of
diffusers-modular/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024.
This repository contains the MiniMax-H3 diffusion transformer only. It does not include the text encoder, tokenizer, VAE, or the rest of the MiniMax-H3 pipeline.
The BF16 checkpoint is a state-dict/layout conversion of the immediate source checkpoint. The two INT8 files are derived native-ComfyUI mixed-precision variants of that converted checkpoint; no training, fine-tuning, additional pruning, or learned-weight adaptation was performed.
Available checkpoints
| File | Variant | Notes |
|---|---|---|
MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-comfy.safetensors |
BF16 | Native ComfyUI conversion; 40,235,519,464 bytes (~37.472 GiB); 534 container tensors |
MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-comfy-int8-fc2bf16.safetensors |
INT8 / BF16 mixed | Native int8_tensorwise; 150 core Linear layers quantized; all 50 MLP fc2 weights remain BF16; 832 tensors |
MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-comfy-int8-convrot-fc2bf16.safetensors |
INT8 ConvRot / BF16 mixed | Same 150-layer policy with ConvRot groupsize 256; all 50 MLP fc2 weights remain BF16; 832 tensors |
BF16 full-file SHA-256:
78b88298e241231b3bd95d752abde711efc9dd6517669a8a934faeb70baf6a98
BF16 tensor-data-region SHA-256:
d5593b33e9920d038241697e871955e6ad118dce08ff49b6bccddfb85e9c06a0
INT8 quantization policy
Both quantized checkpoints use ComfyUI's native per-layer .comfy_quant format and TensorWiseINT8Layout. No custom quantized-model loader is required.
The following weights are quantized in every one of the 50 main transformer blocks:
blocks.N.attn.qkv_proj.weight
blocks.N.attn.out_proj.weight
blocks.N.mlp.fc1.weight
That is 150 quantized Linear layers total.
The following remains BF16:
blocks.N.mlp.fc2.weight
All smaller/sensitive tensors remain in their source precision, including the pruned AdaLN table and projections, final-layer projections, norms, patch/text projections, and token refiner.
Regular INT8 uses tensor-wise scaling:
format: int8_tensorwise
per_channel: false
convrot: false
INT8 ConvRot uses:
format: int8_tensorwise
per_channel: true
convrot: true
convrot_groupsize: 256
The two Diffusers-only auxiliaries adaln_basis and adaln_mean, which are retained in the repaired BF16 artifact but unused by native ComfyUI inference, are omitted from both quantized files.
Why fc2 remains BF16
A fully quantized 200-layer test variant also quantized blocks.N.mlp.fc2.weight. In large MiniMax-H3 sequences, ComfyUI's fused linear_input_act(..., "swiglu") path sends quantized fc2 through comfy_kitchen.int8_linear, which dynamically quantizes the full post-SwiGLU activation matrix. In the tested workflow this produced a very large contiguous INT8 scratch allocation and caused OOM / hard WSL failures.
Keeping fc2 in BF16 avoids that path while retaining INT8 for the other three heavy Linear weights per block. The fully quantized 200-layer experimental files are not the release variants.
Provenance
| Layer | Repository / revision |
|---|---|
| Original model | MiniMaxAI/MiniMax-H3 |
| ComfyUI pruned-model lineage | Comfy-Org/MiniMax-H3 |
| Immediate source | diffusers-modular/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024 |
| Immediate-source revision | c7d26373ecb070f1f1dc8811609d56d10d45d690 |
| Native conversion / quantized packaging | xmarre |
See PROVENANCE.md for the BF16 conversion record.
Pruned MiniMax-H3 timestep conditioning
This checkpoint uses MiniMax-H3's pruned AdaLN-curve representation.
The native model contains:
adaln_t_table [1025, 8] F32
and intentionally does not contain the full-model timestep MLP:
time_embedder.proj_in.*
time_embedder.proj_out.*
The Diffusers source stores the precomputed curve table as time_embedder.table; the native conversion restores it as adaln_t_table.
All 50 transformer-block AdaLN projections and the final-layer AdaLN projection consume the 8-wide curve basis.
Critical folded-bias conversion
The pruned Diffusers checkpoint stores required constant AdaLN modulation terms as separate folded_bias tensors. They are converted to the native ComfyUI bias parameters:
transformer_blocks.N.adaln_proj.folded_bias
-> blocks.N.adaln_proj.linear.bias N = 0..49
norm_out.folded_bias
-> final_layer.adaln_proj.linear.bias
All 51 native constant biases are F32. The learned AdaLN weights preserve their source dtype, BF16 in this checkpoint.
Other structural conversion steps
The BF16 conversion also:
- removes stale Diffusers
configsafetensors metadata; - recombines 52 Diffusers Q/K/V groups into native
qkv_projtensors; - restores the native SwiGLU half ordering for 52 feed-forward tensors;
- reconstructs the non-learned
rope.inv_freqbuffer; - preserves learned-weight dtypes.
The repaired BF16 artifact retains adaln_basis and adaln_mean; native ComfyUI does not consume them. A clean fresh conversion, and both quantized derivatives, omit those auxiliaries.
Validation
The BF16 conversion passed structural validation for the pruned H3 architecture, including:
adaln_t_table [1025, 8] F32: OK
full timestep MLP tensors: absent
all 50 block AdaLN input widths = 8: OK
final-layer AdaLN input width = 8: OK
all 51 native AdaLN biases: present / F32
Diffusers folded_bias keys: absent
stale config metadata: absent
required native MiniMax-H3 keys: present
rope.inv_freq [16] F32: OK
safetensors.safe_open: OK
Both release INT8 variants were additionally validated through full native ComfyUI execution. The tested workflow exercised:
model_type FLOW_AV detection
H3 Continuum main sampling pass 1
H3 Continuum continuation sampling pass 2
Spectrum H3 actual/forecast execution
large 3D latent refine pass
video VAE decode
audio VAE decode
final Continuum assembly
video combine
Both the regular INT8/BF16 and INT8-ConvRot/BF16 variants completed the tested end-to-end workflow without the failures seen when fc2 was quantized.
This is functional compatibility validation for the tested ComfyUI path, not a quantitative image/video quality benchmark.
ComfyUI usage
Place the selected checkpoint in:
ComfyUI/models/diffusion_models/
Then select that exact file in a diffusion-model loader that delegates to ComfyUI's native diffusion-model loading path. The conversion was tested around DiffusionModelLoaderKJ from xmarre/ComfyUI-KJNodes.
For the INT8 files:
weight_dtype: default
compute_dtype: default or bf16
Do not force an additional FP8 weight cast on top of the native INT8 checkpoint.
Use the appropriate MiniMax-H3 text encoder and VAE separately.
Included conversion / repair tools
tools/convert_h3_diffusers_to_comfy_v3.py— corrected Diffusers -> native ComfyUI converter.tools/repair_h3_pruned_comfy_header.py— repairs the pruned timestep-table mapping and removes stale Diffusersconfigmetadata from an earlier conversion.tools/repair_h3_pruned_folded_biases.py— repairs the 51 pruned AdaLN folded-bias names in an earlier converted artifact without rewriting the tensor payload.
Scope
The BF16 file is a native ComfyUI state-dict conversion and packaging of the immediate source checkpoint.
The INT8 files are post-conversion native-ComfyUI quantized derivatives of that BF16 checkpoint. They do not add trained capabilities and are not fine-tunes or additional pruning passes.
License and use restrictions
MiniMax-H3 and derivatives are governed by the MiniMax H3 Community License Agreement. A copy is provided in LICENSE, with the distribution notice in NOTICE.
The upstream license contains territorial, use, distribution, commercial, and acceptable-use restrictions. Review the complete license before using or distributing these weights. This repository does not grant rights beyond the upstream license.
Attribution
- Original MiniMax-H3: MiniMax / Nanonoble Pte. Ltd.
- ComfyUI pruned-model lineage: Comfy-Org
- Immediate Diffusers source: diffusers-modular
- Native ComfyUI conversion and quantized packaging: xmarre, 2026
Related repositories
- Downloads last month
- 429