MiniMax H3 Hybrid FL2VA + REF2VA — GGUF (Q4_0 & Q5_0)
Community GGUF Q4_0 and Q5_0 quantizations of the b25-49 hybrid MiniMax H3 checkpoint.
These GGUFs are derived from:
smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
Specifically:
minimax_h3_hybrid_fl2va_ref2va_b25-49.safetensors
The original hybrid checkpoint combines the FL2VA and REF2VA MiniMax H3 variants, using FL2VA as the base model while using REF2VA adaln_proj weights for transformer blocks 25–49.
This repository provides GGUF Q4_0 and Q5_0 versions intended for use with ComfyUI and other GGUF-compatible runtimes.
Model Information
| Property | Value |
|---|---|
| Architecture | minimax_h3 |
| Variant | b25-49 |
| Quantization | Q4_0 / Q5_0 |
| GGUF version | 3 |
| Tensor count | 532 |
| Tensor elements | 20,111,438,744 |
| GGUF file size (Q4_0) | 10.6 GB (11,387,224,800 bytes) |
| GGUF file size (Q5_0) | 12.9 GB (13,895,623,392 bytes) |
| Format | GGUF |
| Primary runtime | ComfyUI + ComfyUI-GGUF |
Source Model
Original hybrid model:
smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
The original repository contains multiple FL2VA + REF2VA hybrid variants.
These GGUFs were created from the:
b25-49
variant.
Quantization
The source checkpoint was converted to GGUF and quantized to:
Q4_0 and Q5_0
using the tools from:
molbal/ComfyUI-GGUF
https://github.com/molbal/ComfyUI-GGUF
The resulting GGUFs contain:
- 532 tensors
- 20,111,438,744 tensor elements
- 213 quantized tensors (Q4_0 or Q5_0 depending on the file)
- 319 tensors retained in higher precision formats (F16/F32)
- GGUF architecture:
minimax_h3 - GGUF version: 3
The large transformer weight matrices are quantized to Q4_0/Q5_0.
Smaller and/or sensitive tensors, including normalization, bias and modulation-related tensors, remain in F16/F32 to preserve quality.
These are quantizations of the original checkpoint. No additional training or fine-tuning was performed during the conversion.
GGUF Files
Choose the file that best fits your VRAM and quality requirements:
Q4_0 Version (Smaller size, lower VRAM usage)
minimax_h3_hybrid_fl2va_ref2va_b25-49-Q4_0.gguf
Size: 10.6 GB (11,387,224,800 bytes)
Q5_0 Version (Better precision, slightly larger file size)
minimax_h3_hybrid_fl2va_ref2va_b25-49-Q5_0.gguf
Size: 12.9 GB (13,895,623,392 bytes)
ComfyUI Compatibility
These GGUFs were created for use with:
molbal/ComfyUI-GGUF
https://github.com/molbal/ComfyUI-GGUF
The models have been loaded in ComfyUI using the GGUF workflow.
FL2VA Testing
Q4_0: Tested successfully. Q5_0: Untested.
The Q4_0 GGUF was loaded in ComfyUI and successfully generated video using an FL2VA workflow. The Q5_0 version is provided for users with higher VRAM capacity but remains untested in this environment.
REF2VA Testing
Q4_0: Tested successfully. Q5_0: Untested.
The Q4_0 GGUF was loaded in ComfyUI and successfully generated video using a REF2VA reference-conditioning workflow. The REF2VA workflow completed successfully to 100% and produced a video output.
Installation
Install ComfyUI-GGUF:
https://github.com/molbal/ComfyUI-GGUF
Follow the installation instructions provided by the repository.
Place the downloaded GGUF model in the appropriate ComfyUI model directory used by your GGUF workflow.
For example:
ComfyUI/
└── models/
└── unet/
├── minimax_h3_hybrid_fl2va_ref2va_b25-49-Q4_0.gguf
└── minimax_h3_hybrid_fl2va_ref2va_b25-49-Q5_0.gguf
The exact location may depend on the ComfyUI workflow and loader being used.
Hardware
Actual VRAM requirements depend on the workflow and generation settings, including:
- resolution
- number of frames
- reference inputs
- text encoder
- VAE
- GPU offloading
- ComfyUI configuration
- CUDA/PyTorch configuration
The GGUF Q4_0/Q5_0 formats substantially reduce the model weight size compared with the original checkpoint.
Testing Environment: The Q4_0 version has been tested on a system equipped with:
- NVIDIA RTX 5060 Ti 16 GB
⚠️ Important Note on Q5_0: The Q5_0 version has not been tested on this hardware. Due to its larger file size (12.9 GB), it may exceed the 16 GB VRAM limit during generation depending on your workflow settings, potentially resulting in Out Of Memory (OOM) errors.
Performance and VRAM requirements may vary significantly depending on the workflow.
Original Model and Credits
MiniMax
Original MiniMax H3 model and architecture.
smhfacct
Creator of the FL2VA + REF2VA hybrid MiniMax H3 checkpoints.
Source:
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
molbal
Creator of the ComfyUI-GGUF implementation used for GGUF support and conversion.
Repository:
https://github.com/molbal/ComfyUI-GGUF
Disclaimer
This is a community-created GGUF quantization.
It is not an official MiniMax release.
The underlying weights originate from:
smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
No additional training or fine-tuning was performed during the GGUF conversion.
Please refer to the original model and MiniMax licensing terms before using or redistributing this model.
SHA256
Q4_0: 6b14de64d26a3e67316c31b11198eb01bee03cabddbd16cb6f55b65c7b70943c
Q5_0: 21db216adde8836164ec75211f874c45e3d4df36b6a052e1e0b703dff6554aef
Status
| Component | Status |
|---|---|
| Source checkpoint | ✅ b25-49 |
| GGUF conversion | ✅ Complete |
| Q4_0 & Q5_0 quantization | ✅ Complete |
| GGUF header verification | ✅ Passed |
| Tensor inspection | ✅ Passed |
| FL2VA | ✅ Q4_0 Tested / ⚠️ Q5_0 Untested |
| REF2VA | ✅ Q4_0 Tested / ⚠️ Q5_0 Untested |
| Video output | ✅ Confirmed (Q4_0) |