Papina/MiniMax-H3-ref2va-int8

MiniMax-H3 (ref2va, int8)

An int8 weight-only quantized derivative of MiniMaxAI/MiniMax-H3's ref2va workflow, for running on GPUs with less VRAM than the original bf16 checkpoint needs (~130 GB across the transformer and text encoder alone).

What's different from the original

Only the two large components are modified, using torchao Int8WeightOnlyConfig weight-only quantization, converted directly from the original bf16 weights with no retraining or calibration:

• transformer_ref/ — quantized (originally ~67.3 GB bf16)
• text_encoder/ (Qwen3-VL conditioner) — quantized (originally ~59.3 GB bf16)

Everything else this pipeline needs — vae/, audio_vae/, tokenizer/, processor/, scheduler/, audio_scheduler/ — is loaded directly from the original MiniMaxAI/MiniMax-H3 repository, unmodified, via this repo's modular_model_index.json.

This repo only carries the ref2va workflow. For t2va/fl2va, use the original repository's transformer/.

Usage

import torch
from diffusers import ComponentsManager, ModularPipeline

manager = ComponentsManager()
manager.enable_auto_cpu_offload(device="cuda")

pipe = ModularPipeline.from_pretrained(
    "Papina/MiniMax-H3-ref2va-int8", workflow="ref2va", components_manager=manager
)
pipe.load_components(dtype=torch.bfloat16)

See the original model card for the full ref2va usage guide (references, num_frames, etc.) — inference works identically, just with a smaller memory footprint.

License

Distributed under the same terms as the original model — see LICENSE and NOTICE. This is a Model Derivative under the MiniMax H3 Community License Agreement.