apex-flash-1-abliterated-NVFP4

Weight-only NVFP4 quantization of cantina-security/apex-flash-1-abliterated (321B MoE, GLM-5.3-Flash architecture). ~195 GB, vs ~643 GB BF16 / ~328 GB FP8 (FP8 version).

Format

Same layout as dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4 (modelopt quant_algo: NVFP4):

Only the routed MoE experts (mlp.experts.*.{gate,up,down}_proj) are quantized: E2M1 values packed two per byte (U8), FP8 E4M3 weight_scale per 16 input elements, FP32 per-tensor weight_scale_2 (= amax / (6 * 448)).
Weight-only: no input_scale, input_activations: null. Use a W4A16 MoE path (e.g. vLLM Marlin / b12x W4A16), not native W4A4 kernels.
Everything else (attention, KDA, shared experts, dense MLPs, routers, MTP, vision) is BF16, unchanged from the source.
Round-to-nearest, no calibration. Expert tensor relative error ~9% (typical for NVFP4); passthrough tensors are bit-identical to the source.

From the upstream card: this abliterated variant has not undergone a separate evaluation; it is intended for authorized security research.
This quantization has not been separately benchmarked. Not affiliated with Cantina Security or Z.AI.