Qwen3.6-27B-Fable-Fusion-711-Heretic2 — GGUF

GGUF quants of Assiduous42/Qwen3.6-27B-Fable-Fusion-711-Heretic2 — a terminal Heretic ablation of DavidAU's Fable-Fusion-711, targeted at creative-writing refusals rather than AdvBench-style harm refusals.

See the source repo for method, prompt-set design, and honest caveats on the refusal numbers.

Files

No Q5_1 tensors in any quant — see the note at the bottom on why that matters.

Measured on a single RTX 4090 (24 GB)

FF-Heretic2-Q4_K_S in LM Studio, 98304 context, Q4_0 K and V cache:

⚠️ Patched metadata — read this if you compare against the original

The base model is a -MTP build carrying an extra multi-token-prediction layer (blk.64). transformers does not model that head, so it was dropped during the LoRA merge — these weights have 64 layers, not 65.

The converter still read nextn_predict_layers: 1 and wrote block_count: 65, producing files that failed to load with:

check_tensor_dims: tensor 'blk.64.attn_norm.weight' not found

Both keys were corrected in-place in every quant here:

Both are UINT32, so this is a same-width byte write with no offset changes — the tensor data is untouched.

Consequence: MTP speculative decoding is not available. If you were getting "draft tokens accepted" percentages from the original, you will not here.

A note on Q5_1

Worth checking on any GGUF before you trust it: if a quant contains Q5_1 tensors in the MoE/expert path, llama.cpp may fall back to CPU for those matmuls and the model will load almost entirely into system RAM despite ample free VRAM — silently, with no error. Symptoms are ~100% reported GPU utilisation at very low power draw.

You can check the tensor-type histogram from the GGUF header alone, without downloading the whole file, using an HTTP range request for the first few MB.

These quants are clean:

Q4_K 485, F32 353, Q5_K 12, Q6_K 1

Suggested settings

• Context 98304 with Q4_0 K/V cache fits comfortably on 24 GB at Q4_K_S.
• This is a thinking model. Short max_tokens will be consumed entirely by the reasoning channel and you will get empty content. Either allow 600+ tokens, or disable thinking: chat_template_kwargs: {"enable_thinking": false}

License

Apache 2.0, inherited. Derivative work.