apex-flash-1-heretic-v2-NVFP4

NVFP4 quantization of pqhaz/apex-flash-1-heretic-v2: modelopt weight-only layout, routed experts only (gate/up/down projections, block size 16), everything else left in BF16. Tensor names, shard assignment and quantization_config are the same as in pqhaz/apex-flash-1-abliterated-NVFP4, so it is a drop-in replacement for that checkpoint.

Produced with nvfp4.py (included). The converter was checked against a shard of the earlier NVFP4 repo, converting the same source weights: packed weights and both scales came out bit-identical on the 25 expert tensors sampled.

See the BF16 repository for what was changed and how it was measured; ablation.json holds the exact parameter set. This quantization itself has not been evaluated.

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