10Eros-Max H3 — INT8 ConvRot
Native ComfyUI INT8 ConvRot conversions of TenStrip/10Eros-Max.
Available models
TURBO Hybrid Beta3
- Full INT8 ConvRot — 200 quantized layers.
- INT8 ConvRot with skip edges — 184 quantized layers; blocks 0, 1, 48, and 49 remain at BF16.
See SKIP_EDGES.md for a detailed explanation of the skip-edges variants.
Beta2
Existing Beta2 INT8 ConvRot variants remain available:
10Eros_Max_h3_TURBO_ref2va_beta2_int8_convrot.safetensors10Eros_Max_h3_TURBO_ref2va_beta2_int8_convrot_skip_edges.safetensors10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot.safetensors10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot_skip_edges.safetensors10Eros_Max_h3_ref2va_beta2_pruned_int8_convrot.safetensors10Eros_Max_h3_ref2va_beta2_pruned_int8_convrot_skip_edges.safetensors
Conversion format
- Native ComfyUI
int8_tensorwise - Per-channel absmax quantization
- ConvRot enabled
- Groupsize: 256
- Source precision: BF16
The skip-edges variants preserve the model’s boundary blocks at higher precision and are the more conservative quality-focused option.
Usage
Place the selected file in:
ComfyUI/models/diffusion_models
Use a ComfyUI version with native INT8 ConvRot support.
License
This model follows the MiniMax H3 Community License Agreement. Review the license terms before use or redistribution.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support