apex-flash-1-heretic-v2-FP8

FP8 (e4m3, 128x128 block scales) quantization of pqhaz/apex-flash-1-heretic-v2, in the same tensor layout and quantization_config as the official zai-org/GLM-5.3-Flash release. Produced with convert.py (included), the same script as pqhaz/apex-flash-1-abliterated-FP8.

See the BF16 repository for what was changed and how it was measured; ablation.json holds the exact parameter set. The numbers there were measured on the BF16 weights, not on this quantization.

Intended for authorized security research in environments the researcher owns or has permission to test.