Booper-Big-Chat-INT8

This is the real INT8-on-disk export of ProCreations/Booper-Big-Chat. All matrix weights, including the 3-D MoE expert tensors, use symmetric per-output-channel INT8; norms remain BF16. The weight artifact is 164.7 MB, a 2.00× reduction from the BF16 safetensors file.

Because native Transformers quantizers do not currently wrap Mixtral's 3-D expert tensors, load_int8.py is included. It reconstructs the standard Mixtral model in BF16 for maximum compatibility while retaining a compact downloadable INT8 artifact:

from huggingface_hub import snapshot_download
import sys

path = snapshot_download("ProCreations/Booper-Big-Chat-INT8")
sys.path.insert(0, path)
from load_int8 import load_model
model = load_model(path, device="cuda")