---

tags:

• text-generation-inference
• transformers
• qwen3.6
• heretic
• uncensored
• decensored
• abliterated
• reproducible
• llama-cpp
• gguf-my-repo

license: apache-2.0

base_model:

• darkc0de/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-heretic

datasets:

• TeichAI/claude-4.5-opus-high-reasoning-250x
• TeichAI/Claude-Opus-4.6-Reasoning-887x

---

Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-heretic — GGUF

GGUF quantizations of darkc0de/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-heretic

Model Overview

| Property | Value |

|---|---|

| Base Model | Qwen/Qwen3.6-27B |

| Architecture | Qwen3.5ForConditionalGeneration |

| Parameters | 27B (dense) |

| Context Length | 262,144 tokens (256K) |

| License | Apache-2.0 |

Training Pipeline

Step 1: Claude Opus Reasoning Distillation

Fine-tuned on Claude Opus reasoning datasets using Unsloth + HuggingFace TRL (2x training speedup):

• TeichAI/Claude-Opus-4.6-Reasoning-887x
• TeichAI/claude-4.5-opus-high-reasoning-250x

Step 2: Abliteration (Uncensoring)

Applied using Heretic v1.3.0. Reproducible via reproduce/ directory in the original repo.

Abliteration Parameters:

| Parameter | Value |

|---|---|

| direction_index | 41.77 |

| attn.o_proj.max_weight | 1.22 |

| attn.o_proj.max_weight_position | 51.32 |

| attn.o_proj.min_weight | 1.21 |

| attn.o_proj.min_weight_distance | 32.79 |

| mlp.down_proj.max_weight | 1.44 |

| mlp.down_proj.max_weight_position | 42.02 |

| mlp.down_proj.min_weight | 0.13 |

| mlp.down_proj.min_weight_distance | 37.15 |

Quality Metrics

KL Divergence & Refusal Rate

| Metric | Value |

|---|---|

| KL Divergence (vs base) | 0.0774 |

| Refusals | 8/100 (down from 98/100) |

Benchmarks (mxfp8 precision)

| Benchmark | Base Qwen3.6-27B | This Model | Delta |

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

| ARC | 0.647 | 0.665 | +1.8% |

| ARC-E | 0.803 | 0.831 | +2.8% |

| BoolQ | 0.910 | 0.910 | 0.0% |

| HellaSwag | 0.773 | 0.790 | +1.7% |

| OpenBookQA | 0.450 | 0.456 | +0.6% |

| PIQA | 0.806 | 0.813 | +0.7% |

| WinoGrande | 0.742 | 0.772 | +3.0% |

Available Quantizations

| File | Size | BPW |

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

| *-f16.gguf | ~50 GB | 16.0 |

| *-Q1_0.gguf | ~4 GB | 1.1 |

| *-Q2_K.gguf | ~10 GB | 2.5 |

| *-Q3_K_S.gguf | ~11 GB | 3.0 |

| *-Q3_K_M.gguf | ~12 GB | 3.5 |

| *-Q3_K_L.gguf | ~13 GB | 3.8 |

| *-Q4_0.gguf | ~14 GB | 4.0 |

| *-Q4_1.gguf | ~16 GB | 4.8 |

| *-Q4_K_S.gguf | ~15 GB | 4.0 |

| *-Q4_K_M.gguf | ~15 GB | 4.5 |

| *-Q5_0.gguf | ~17 GB | 5.0 |

| *-Q5_1.gguf | ~19 GB | 5.7 |

| *-Q5_K_S.gguf | ~17 GB | 5.0 |

| *-Q5_K_M.gguf | ~18 GB | 5.5 |

| *-Q6_K.gguf | ~21 GB | 6.0 |

| *-Q8_0.gguf | ~27 GB | 8.0 |

Quantization Details

• f16: Full float16 precision, lossless conversion from safetensors
• Q1_0: Extreme compression, very low quality
• Q2_K: Smallest practical size, lower quality
• Q3_K_S / Q3_K_M / Q3_K_L: Q3 variants with increasing quality
• Q4_0 / Q4_1: Legacy formats, compatible with older tools
• Q4_K_S / Q4_K_M: K-quant mixtures with balanced quality/size
• Q5_0 / Q5_1: Legacy formats, compatible with older tools
• Q5_K_S / Q5_K_M: K-quant mixtures with high quality
• Q6_K: Highest quality K-quant
• Q8_0: Near-lossless, 8-bit integer quantization

System Requirements

| Quant | VRAM |

|---|---|

| f16 | ~50 GB |

| Q1_0 | ~4 GB |

| Q2_K | ~10 GB |

| Q3_K_S | ~11 GB |

| Q3_K_M | ~12 GB |

| Q3_K_L | ~13 GB |

| Q4_0 | ~14 GB |

| Q4_1 | ~16 GB |

| Q4_K_S | ~15 GB |

| Q4_K_M | ~15 GB |

| Q5_0 | ~17 GB |

| Q5_1 | ~19 GB |

| Q5_K_S | ~17 GB |

| Q5_K_M | ~18 GB |

| Q6_K | ~21 GB |

| Q8_0 | ~27 GB |

Recommended Settings

Temperature: 0.7
Top-p: 0.9
Top-k: 40
Repeat penalty: 1.1

Links

• Original model: darkc0de/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-heretic
• Base model: Qwen/Qwen3.6-27B
• Distillation source: TeichAI/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2
• Heretic tool: Heretic v1.3.0