molbal/MiniMax-H3-Turbo-GGUF

MiniMax-H3-Turbo (FL2VA 4-step v1.0 768p GGUF Quantizations)

This repository provides GGUF quantized versions and ComfyUI workflows for the FL2VA Turbo 4-step v1.0 768p variant of MiniMax-H3-Turbo. MiniMax H3 is an omni-modal generative system supporting unified text, image, video, and audio generation with native stereo audio.

The Turbo distillation compresses inference down to 4 steps (NFE) at 768p resolution with shifted schedules.

Full Merged Checkpoints: These GGUF files contain full merged weights (the base model + distilled turbo weights merged directly into a standalone model checkpoint), not separate LoRA adapters. You load them directly as diffusion models in ComfyUI without needing additional base weights or LoRA loaders.

Model Specifications

Parameter	Specification
Distilled Variant	FL2VA Turbo 4-step v1.0 768p
Tasks	FL2VA / I2VA / T2VA (First/Last Frame & Text to Audio-Video)
Training Resolution	768p (1344×768)
Training Shifts (Video / Audio)	6 / 3
Distillation Steps (NFE)	4
Recommended Inference Steps (NFE)	4
Frame Rate	24 FPS
Audio Output	32 kHz stereo

Repository Contents

GGUF Models

• minimax_h3_fl2v_turbo_4step_v1.0_768p_Q4_0.gguf (11.4 GB)
• minimax_h3_fl2v_turbo_4step_v1.0_768p_Q8_0.gguf (21.4 GB)
• minimax_h3_fl2v_turbo_4step_v1.0_768p_Q8_CR.gguf (20.2 GB)

Quantization Formats

Q8_CR is specifically created to work with the molbal/ComfyUI-GGUF ComfyUI custom nodes.

Format	Quality	Notes
Q4_0	Lower	Standard 4-bit GGUF quantization.
Q8_0	High	Standard 8-bit GGUF quantization.
Q8_CR	High (≈ Q8_0)	Q8_0 quality with INT8 ConvRot weights.

ComfyUI Usage & Node Requirements

To run these GGUF models in ComfyUI, install the following custom node packages:

1. GGUF Loader Node (one of the following):
   • ComfyUI-GGUF (GitHub) or install via ComfyUI Node Registry: comfyui-gguf-reboot
   • Used to load the quantized .gguf UNet/diffusion models.
2. MiniMax H3 Turbo Sampler (one of the following):
   • ComfyUI-MiniMax-H3-Turbo (GitHub) or install via ComfyUI Node Registry: comfyui-minimax-h3-turbo
   • Provides the dedicated Turbo sampling schedules and shift calculations for 4-step generation.

Model Placement

Place the downloaded GGUF files in your ComfyUI models directory:

ComfyUI/models/diffusion_models/minimax_h3_fl2v_turbo_4step_v1.0_768p_Q8_CR.gguf

(or ComfyUI/models/unet/ depending on your GGUF loader configuration)

Shift & Sampling Schedule

For NFE = 4, video shift = 6, and audio shift = 3:

• Define unshifted evaluation points: $q_i = \frac{4 - i}{4}$ for $i = 0, 1, 2, 3 \rightarrow [1.0, 0.75, 0.5, 0.25]$
• Recommended inference step count: 4 steps.

Supported Input Modes (FL2VA)

• Text-to-Audio-Video (T2VA): No image input connected.
• Image-to-Audio-Video (I2VA): One image connected to first_frame.
• First-and-Last-Frame-to-Audio-Video (FL2VA): Two images connected to first_frame and last_frame.

ComfyUI Workflows

Example workflows are provided in the workflows/ directory:

• workflows/minimax_h3_t2v-turbo-gguf.json — Text-to-Audio-Video (T2VA) Turbo workflow.
• workflows/minimax_h3_i2v-turbo-gguf.json — Image-to-Audio-Video (I2VA / FL2VA) Turbo workflow.

License

MiniMax H3 is released under the MiniMax H3 Community License Agreement.