DmitryDB/MiniMax-H3-ComfyUI-Quants

MiniMax-H3 Stock ComfyUI Quants

Community FL2VA and Ref2VA diffusion-transformer checkpoints for MiniMaxAI/MiniMax-H3. All files in this repository retain all 50 transformer blocks and use the stock ComfyUI fused-QKV and time-table layout. No custom node or ComfyUI core patch is required.

These are community conversions, not official MiniMax or ComfyOrg releases.

Naming

• No runtime marker in the filename means stock ComfyUI compatible.
• FL2VA is text/first-frame/last-frame-to-audio-video generation.
• Ref2VA is reference-image/video/audio-to-audio-video generation.
• Quantized tensor counts, retained BF16 islands, GPU class, and expected memory class are documented here instead of being encoded in filenames.
• The patch-required dynamic-time, separate-QKV editions use the explicit DT-sQKV marker and live in the separate MiniMax-H3-DynTime-sQKV repository.

Choose a checkpoint

Download one FL2VA or Ref2VA checkpoint from the same profile row.

Profile	Direct downloads	File size	GPU class and quant layout
INT8 ConvRot HQ	FL2VA · Ref2VA	21.908 GiB	32 GB+ · RTX 30/40. 145 INT8 ConvRot + 63 BF16 semantic matrices. Largest BF16 island. A 24 GB RTX 4090 loader test offloaded about 0.955 GiB.
INT8 ConvRot	FL2VA · Ref2VA	20.940 GiB	24 GB · RTX 30/40. 170 INT8 ConvRot + 38 BF16 semantic matrices. Fully resident in the RTX 4090 loader test.
INT8 ConvRot Lite	FL2VA · Ref2VA	20.330 GiB	24 GB · RTX 30/40. 185 INT8 ConvRot + 23 BF16 semantic matrices. Leaves more memory for the rest of the workflow.
W8/W4 ConvRot	FL2VA · Ref2VA	13.565 GiB	16 GB · RTX 30/40. 86 W8 + 114 W4 main matrices; the eight token-refiner matrices remain BF16.
W4 ConvRot	FL2VA · Ref2VA	10.067 GiB	12 GB · RTX 30/40. 200 W4 main matrices + 8 INT8 token-refiner matrices.
W4 ConvRot Offload	FL2VA · Ref2VA	9.708 GiB	8 GB + CPU offload · RTX 30/40. All 208 main and token-refiner matrices use W4.
NVFP4 HQ	FL2VA · Ref2VA	13.597 GiB	16–24 GB · RTX 50/Blackwell. 170 NVFP4 + 30 BF16 main matrices; the eight token-refiner matrices remain BF16.
NVFP4	FL2VA · Ref2VA	10.862 GiB	8–12 GB · RTX 50/Blackwell. All 208 main and token-refiner matrices use block-scaled NVFP4.

Checkpoint size is not full-workflow peak VRAM. Resolution, frame count, attention backend, text encoder, VAEs, and ComfyUI offload settings also affect memory use. RTX 50 classifications are architecture-based; no full generation run was performed on an RTX 5090. NVFP4 here is block-scaled NVFP4, not AWQ.

Measured RTX 4090 loader results

FL2VA and Ref2VA were tested independently through stock ComfyUI. Each test also executed a real quantized INT8 projection.

Checkpoint	Loaded weights	Peak reserved	Free after load	Result
MiniMax-H3_*VA-INT8-ConvRot-Lite.safetensors	100%	20.424 GiB	2.072 GiB	PASS
MiniMax-H3_*VA-INT8-ConvRot.safetensors	100%	21.025 GiB	1.471 GiB	PASS
MiniMax-H3_*VA-INT8-ConvRot-HQ.safetensors	95.6%	21.002 GiB	1.494 GiB	PASS; about 0.955 GiB offloaded

These are loader/kernel measurements, not complete prompt-to-decoded-video VRAM peaks.

Quantization and preserved components

All 16 diffusion checkpoints:

• retain all 50 transformer blocks;
• use the fused qkv_proj = cat(Q,K,V) layout expected by stock ComfyUI;
• use a rank-16 FP32, 4,097-point time table;
• retain 51 independent FP32 AdaLN projections;
• keep norms, conditioning projections, patch projections, output heads, and other small or sensitive tensors in source precision;
• load without a custom loader or core patch in tested ComfyUI commit 14b05228.

INT8, W8, and W4 weights use ConvRot/Hadamard rotation with group size 256, per-row FP32 scales, and deterministic scale search.

INT8 profile	BF16 attention-output blocks	BF16 MLP fc2 blocks	Token refiner
INT8-ConvRot-Lite	0, 1, 2, 3, 4, 5, 6, 7, 9, 15, 19, 38, 45, 49	49	Eight BF16 matrices
INT8-ConvRot	0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 19, 20, 27, 38, 43, 44, 45, 46, 47, 49	39, 45, 49	Eight BF16 matrices
INT8-ConvRot-HQ	all blocks 0–49	29, 39, 44, 45, 49	Eight BF16 matrices

Stock layout versus DT-sQKV

Feature	This repository	DT-sQKV repository
Attention storage	Fused qkv_proj	Separate q_proj, k_proj, v_proj
Attention execution	One fused call	Three projection calls
Original FP32 time_embedder	Replaced by measured time table	Retained and evaluated at runtime
adaln_t_table	FP32 [4097,16]	Absent
Per-block AdaLN	51 independent FP32 rank-16 projections	51 independent FP32 rank-16 projections
ComfyUI	Stock	Core patch required

The time table does not remove timestep conditioning. It interpolates a compact representation of the original measured time curve. Maximum measured table interpolation error is below 0.001%; sampled end-to-end AdaLN relative error is approximately 3e-7 to 4e-7 across 19 timesteps.

Validation

Every released checkpoint passed:

1. exact key, shape, dtype, and quantization-inventory checks;
2. sampled reconstruction against its original FL2VA or Ref2VA HF shards;
3. a 19-timestep FP32 AdaLN numerical comparison;
4. complete CPU load as MiniMaxH3Model in clean ComfyUI commit 14b05228;
5. remote byte-size and LFS SHA-256 verification.

Reports under reports/ retain their historical internal profile names so the published validation provenance remains intact. BF16 samples were checked bit-for-bit. A representative INT8 QKV sample has relative L2 error 0.008814. A prompt-to-decoded-video perceptual A/B score has not been measured.

Installation and required components

Place one selected FL2VA or Ref2VA checkpoint in:

ComfyUI/models/diffusion_models/

A complete workflow also needs the separately maintained Qwen3-VL MiniMax-H3 text encoder and these shared VAEs:

File	Role
vae/MiniMax-H3_VideoVAE-FP16.safetensors	Video latent encoder and decoder
vae/MiniMax-H3_AudioVAE-FP32.safetensors	Audio latent encoder and decoder

No text encoder is included in this repository.

License and attribution

Use is subject to the included MiniMax-H3 community license. The base model is by MiniMax. This community conversion is not endorsed by MiniMax or ComfyOrg.