ghost-actual/qwen3.6-27b-apostate-Q4_K_M-GGUF overview

Qwen3.6-27B-Apostate Q4_K_M GGUF

Q4_K_M quantization of heterodoxin/qwen3.6-27b-apostate — a clean weight-edit abliteration of the base Qwen3.6-27B with refusal behavior removed. No finetune, no LoRA, no runtime hooks. Just the refusal direction projected out of the residual stream weights.

Quantization Details

• Method: Q4_K_M (4.92 BPW)
• Tool: llama.cpp (build c818263, Jun 28 2026)
• Source: BF16 safetensors → F16 GGUF → Q4_K_M
• Size: 15.7 GB

MTP Metadata Fix

.attn_norm.weight'`.

The base Apostate model inherits MTP (multi-token prediction) metadata from Qwen3.6-27B (nextn_predict_layers = 1, block_count = 65) but does not include the MTP head weights (blk.64). This causes llama.cpp to error with `missing tensor 'blk.64

This GGUF has been patched:

• qwen35.block_count: 65 → 64
• qwen35.nextn_predict_layers: 1 → 0

If you quantize the Apostate model yourself, you'll need to apply the same fix or llama.cpp will refuse to load it.

Performance

Tested on RTX 3090 Ti with llama.cpp (flash-attn, asymmetric KV q8_0/q4_0, 128K context, parallel 2):

| Metric | Value |

|--------|-------|

| Generation speed | ~41 tok/s |

| Prompt eval | ~115 tok/s |

| VRAM usage | ~22 GB (128K ctx, 2 slots) |

Usage

llama-server \
    -m qwen3.6-27b-apostate-Q4_K_M.gguf \
    -ngl -1 --ctx-size 131072 \
    --flash-attn on \
    --cache-type-k q8_0 --cache-type-v q4_0

Credits

• Abliteration: heterodoxin using Apostate
• Base model: Qwen/Qwen3.6-27B