MiniMax H3 hybrid — mixed INT8 / W4A8 (24 GB)
Single-file ComfyUI diffusion transformer. A compact hybrid checkpoint with mixed native quantization for 24 GB cards.
Put minimax_h3_hybrid_fl2va_ref2va_b45-49_rtx3090_w4a8-int8.safetensors in ComfyUI/models/diffusion_models/ and load with the stock Load Diffusion Model node on a current ComfyUI that supports INT8 ConvRot and W4A8. Use the usual H3 text encoder and video/audio VAEs.
File
| File | minimax_h3_hybrid_fl2va_ref2va_b45-49_rtx3090_w4a8-int8.safetensors |
| Size | ~18 GiB |
| Format | mixed INT8 + W4A8, compact H3 layout |
Sensitive / small tensors stay at higher precision. Main block weights are a mix of INT8 and W4A8. Exact layer map is not documented here.
Text encoder and VAEs are not included.
Original sources
- Base model: MiniMaxAI/MiniMax-H3
- ComfyUI packaging: Comfy-Org/MiniMax-H3
- Hybrid this file was built from: diobrando0/MiniMax-H3-fl2va-ref2va-hybrids-bf16
- Hybrid method origin: smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
- License: MiniMax H3 Community License Agreement — this conversion grants no extra rights.
Model tree for cicalooo/MiniMax-H3-hybrid-b45-49-rtx3090-w4a8-int8
Merge model
this model