Stock-ComfyUI quantized FL2VA checkpoints for the 10Eros_Max fine-tune of
MiniMaxAI/MiniMax-H3.
10Eros_Max retains the full MiniMax-H3 architecture and all 50 transformer
blocks. Its fine-tuned weights are confined to the fused QKV projections in
blocks 0–31. These files preserve those fine-tuned QKV weights while retaining
the tested quantization layouts and all remaining tensors from the corresponding
MiniMax-H3 reference quants.
No custom node or ComfyUI core patch is required. These are community
conversions, not official MiniMax or ComfyOrg releases.
32 GB+, or 24 GB with limited offload · RTX 30/40. 145 INT8 ConvRot + 55 BF16 main matrices. All attention-output projections and five sensitive FC2 projections remain BF16.
8–12 GB · RTX 50/Blackwell. All 200 main and eight token-refiner matrices use block-scaled NVFP4. NVFP4 is not AWQ.
Checkpoint size is not full-workflow peak VRAM. Resolution, frame count,
attention backend, text encoder, VAEs, and ComfyUI offload settings also affect
memory use. The memory classes above follow the identical tensor layouts of the
corresponding MiniMax-H3 reference quants; a full prompt-to-decoded-video VRAM
measurement has not yet been performed for 10Eros_Max.
What was preserved
All four checkpoints:
retain all 50 MiniMax-H3 transformer blocks;
use the stock-ComfyUI fused qkv_proj = cat(Q,K,V) representation;
preserve the 10Eros_Max fine-tuned fused-QKV weights in blocks 0–31;
preserve the original/reference QKV weights in blocks 32–49;
use the rank-16 FP32, 4,097-point AdaLN time table;
retain 51 independent FP32 AdaLN projections;
preserve norms, conditioning projections, patch projections, output heads,
and other small or sensitive tensors in the same representation as their
corresponding reference checkpoint.
The time table does not remove timestep conditioning. It interpolates a compact
representation of the original measured time curve.
Quantization layouts
INT8 weights use tensorwise INT8 ConvRot/Hadamard rotation with group size 256,
per-row FP32 scales, and deterministic scale search.
The compact NVFP4 profile uses block-scaled NVFP4 for every main and
token-refiner linear matrix. The NVFP4-HQ profile follows the same sensitive
layer policy as INT8-ConvRot: 170 main matrices use NVFP4, while 30 main
matrices and all eight token-refiner matrices remain BF16. Both NVFP4 profiles
are intended primarily for Blackwell hardware.
The experimental W4, W4-Offload, and mixed W8/W4 layouts are intentionally not
published for 10Eros_Max. Their base-model checkpoints failed perceptual video
tests even though their containers and tensor layouts were structurally valid.
Conversion method
The complete 10Eros_Max BF16 checkpoint was compared tensor-by-tensor with the
local original MiniMax-H3 FL2VA source:
all 535 expected tensors and all 50 blocks were present;
shapes and dtypes matched the stock full-width architecture;
no NaN or Inf values were found;
503 tensors were bit-identical to the original source;
the intentional fine-tune consisted of 32 modified fused-QKV tensors in
blocks 0–31.
Only those 32 QKV tensors were requantized for each runtime format. The INT8
QKV payload was produced once and reused for both INT8 profiles; the NVFP4 QKV
payload was likewise reused for both NVFP4 profiles. Each final checkpoint was
assembled from its corresponding tested reference checkpoint by replacing only
the QKV weight, scale, and quantization-configuration byte ranges for blocks
0–31.
This avoids requantizing or rewriting the semantic content of unchanged layers.
Validation
All four published files passed an independent byte-range audit:
file size, safetensors header, key inventory, shapes, and dtypes match the
corresponding reference checkpoint;
every byte outside the permitted QKV ranges remains identical to the
reference checkpoint;
every replaced QKV/scale/config tensor matches the separately generated
10Eros_Max quantized payload;
the replacement weights differ from the base reference as expected;
the local INT8 and NVFP4 quantizers reproduced the reference quantized
tensors bit-for-bit on an unchanged control block.
No prompt-to-decoded-video perceptual A/B result is claimed yet. The validation
above proves structural and conversion correctness, not subjective fine-tune
quality.
Installation
Place one selected checkpoint in:
ComfyUI/models/diffusion_models/
Then refresh the model list or restart ComfyUI and load it through the standard
MiniMax-H3 FL2VA workflow.
A complete workflow also requires the MiniMax-H3 Qwen3-VL text encoder and the
video/audio VAEs. They are not duplicated here. Shared components and additional
base-model quant profiles are documented in
MiniMax-H3-ComfyUI-Quants.
License and attribution
Use is subject to the included MiniMax-H3 Community License Agreement. The base
model is by MiniMax. The BF16 fine-tune was supplied under the name
10Eros_Max; this repository provides quantized conversions and does not
claim authorship of the fine-tune itself.
ComfyUI and its quantization runtimes are separate upstream projects. This
repository is not endorsed by MiniMax or ComfyOrg.
