CoolFace
Modelpublic

jan1k/Qwen3.8-27B-Uncensored-Genesis-NVFP4

sourceHugging Faceapache-2.0updated 5d agoView on Hugging Face
3likes371downloads
Model Card

Qwen3.8-27B-Uncensored-Genesis — NVFP4 GGUF

NVFP4 quantisation of LuffyTheFox/Qwen3.8-27B-Uncensored-Genesis-GGUF, built with advanced-gguf-quantizer (a llama.cpp fork focused on NVFP4 quantization).

This is the Genesis-repaired version of HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive — Luffy's tensor repair applied to the HauhauCS uncensored base.

Files

v4 — Recommended (inline scales, LM Studio compatible)

FileCalibrationMTPSizeTensors
Qwen3.8-27B-Uncensored-Genesis-NVFP4-v4.ggufnone (data-free)yes (blk.64 preserved)15.76 GB866
Qwen3.8-27B-Uncensored-Genesis-NVFP4-v4-noMTP.ggufnone (data-free)no (blk.64 stripped)15.53 GB851

v4 uses --nvfp4-inline-scales-only to emit native inline UE4M3 scales without separate .scale/.input_scale tensors. Required for LM Studio / Pelican and other runtimes that do not support the extended NVFP4 scale tensor contract.

Imatrix variants will follow after the plain v4 is confirmed working in target runtimes.

Source

Source GGUFQwen3.8-27B-Uncensored-Genesis-Q8_K_P.gguf (31.5 GB)
Architectureqwen35 (dense), 64 layers + 1 MTP/NextN block
Layers48 Gated DeltaNet + 16 gated-attention
general.file_type39 (LLAMA_FTYPE_MOSTLY_NVFP4)
MTP/NextNqwen35.nextn_predict_layers=1 (native, preserved)

Single-step Q8KP → NVFP4 with per-tensor-type protections.

Tensor mix (v4)

typecountnotes
F32360norms, ssm scalars (a/dt/conv1d), nextn norms
F164blk.0 attngate/attnqkv, blk.0/blk.13 ffn_down
NVFP4502bulk weights incl. output.weight, token_embd
total866
separate .scale/.input_scale0inline UE4M3 only

Tensor protection policy

F16 singular-collapse protection:

tensortype
blk.0.attn_gate.weightF16
blk.0.attn_qkv.weightF16
blk.0.ffn_down.weightF16
blk.13.ffn_down.weightF16

F32 architecture-specific protection:

  • —blk.*.attn_norm.weight, blk.*.post_attention_norm.weight
  • —blk.*.attn_q_norm.weight, blk.*.attn_k_norm.weight
  • —blk.*.ssm_norm.weight, blk.*.nextn.*.norm.weight
  • —output_norm.weight
  • —blk.*.ssm_conv1d.weight, blk.*.ssm_dt.bias, blk.*.ssm_a

Forced NVFP4 (do not push lower):

tensortype
blk.0.ssm_out.weightNVFP4
blk.1.attn_gate.weightNVFP4
blk.1.attn_qkv.weightNVFP4

Usage

bash
llama-cli -m Qwen3.8-27B-Uncensored-Genesis-NVFP4-v4.gguf \
  --jinja -c 131072 -ngl 99

For runtimes without FastMTP support, use Qwen3.8-27B-Uncensored-Genesis-NVFP4-v4-noMTP.gguf. The MTP variant enables speculative decoding (--spec-type draft-mtp where supported).

  • —Set K cache and V cache quantization to F16
  • —Vision support requires the mmproj file from the source repository

Hardware

  • —Blackwell (RTX 50xx): native FP4 path, fastest
  • —Ampere (RTX 30xx): NVFP4 inference works via fallback kernels
  • —Quantisation was done CPU-only (Ampere CUDA NVFP4 encoder is unreliable)

Credits