Minimax-H3-Sulphur3-10Erosmax-Merge
a hybrid video generation model based on MiniMax-H3
2 min read
This model is the result of a comprehensive merge of the MiniMax-H3, Sulphur-3, and 10Eros-Max architectures with baked-in LoRA adapters to speed up performance. The build is optimized for video generation in ComfyUI and is available in both full-weight BF16 and quantized INT8 ConvRot formats to save VRAM. The model combines the strengths of the base checkpoints, improving generation quality and speed through built-in distillation additions (LoRA, small auxiliary optimization modules).
Key features
- Complex merge recipe: A 50/50 linear blend of the Sulphur-3 and 10Eros-Max models.
- Baked LoRAs: DMAD and PDMD adapters are integrated into the weights for rapid generation in fewer steps.
- Two formats: Full-size BF16 and native INT8 ConvRot with weight quantization for memory efficiency are available.
- ComfyUI-native: Full compatibility with standard diffusion model loading nodes.
What you can do
- Generate video: Create high-quality clips in ComfyUI leveraging the high speed of baked distillation LoRAs.
- Save VRAM: Run the quantized
int8_convrotversion on GPUs with limited memory. - Integrate into pipelines: Use the standard Load Diffusion Model (UNETLoader) node to hook it into existing workflows.
Example
Place the s3_10ErosMax_dmad085_pdmd0333_bf16.safetensors file or the INT8 version into the ComfyUI/models/diffusion_models/ directory and load it via the standard Load Diffusion Model node (ComfyUI >= v0.27.0 required).
Read more
DetailsBase: MiniMax-H3, Sulphur-3, 10Eros-Max
File formats: BF16 and INT8 ConvRot (int8_tensorwise, group size 256)
Quantization details: Attention layer and MLP are quantized to INT8, sensitive layers (adaln, normalization, rope, etc.) are kept in BF16
License: MiniMax H3 Community License














Comments