10Eros-Max-h3-int8-convrot

10Eros-Max H3 - INT8 ConvRot

INT8 ConvRot conversions of TenStrip/10Eros-Max, originally authored by TenStrip.

The ref2va Beta2 variants use the corrected upstream 10Eros_Max_h3_ref2va_beta2_pruned.safetensors source. Both use ComfyUI's native int8_tensorwise format with per-channel absmax quantization, ConvRot enabled, and a ConvRot groupsize of 256.

Available variants

• Ref2VA Beta2 full INT8 ConvRot - 200 quantized layers. All recipe-selected layers, including edge blocks, use INT8 ConvRot.
• Ref2VA Beta2 INT8 ConvRot with skip edges - 184 quantized layers. Transformer blocks 0, 1, 48, and 49 remain at source precision; the remaining recipe-selected layers use INT8 ConvRot.

The skip-edges variant is the more conservative option for the model's input and output boundaries and uses more storage and VRAM than the full variant.

The existing fl2va Beta2 files remain preserved in this repository:

• 10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot.safetensors
• 10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot_skip_edges.safetensors

Conversion details

• Corrected ref2va Beta2 source: TenStrip/10Eros-Max/10Eros_Max_h3_ref2va_beta2_pruned.safetensors
• Source dtype: BF16
• Quantization: per-channel absmax INT8 tensorwise
• Format marker: int8_tensorwise
• ConvRot: enabled
• ConvRot groupsize: 256
• Ref2VA Beta2 full quantized layers: 200
• Ref2VA Beta2 skip-edges quantized layers: 184
• Skip-edge blocks: 0, 1, 48, and 49 (left at source precision)

The conversion uses the local ref2va INT8 ConvRot checkpoint as the layer and format recipe.

License

This model follows the same license declaration as the original TenStrip/10Eros-Max repository: the MiniMax H3 Community License Agreement.

Please review the license terms, including applicable territory restrictions, use restrictions, redistribution requirements, and notice requirements, before using or distributing this model.

Usage

Place the selected .safetensors file in:

ComfyUI/models/diffusion_models

Load it with a ComfyUI version that supports native INT8 ConvRot models.