Base model credit: Fusion is Kyle Hessling’s reasoning⊕coding merge of Qwen3.6-27B lineage. This repo only adds Heretic abliteration + GGUF packaging. See the Fusion BF16 card for merge method, benchmarks, and intended use.

Files

File	Quant	Approx. size	Notes
`Qwopus3.6-27B-Fusion-heretic-MTP-F16.gguf`	F16	~51 GB	Full-precision GGUF (conversion source)
`Qwopus3.6-27B-Fusion-heretic-MTP-Q8_0.gguf`	Q8_0	~28 GB	High-fidelity; also used as imatrix host
`Qwopus3.6-27B-Fusion-heretic-MTP-Q5_K_M.gguf`	Q5_K_M	~19 GB	Imatrix-guided; strong quality/size
`Qwopus3.6-27B-Fusion-heretic-MTP-Q4_K_M.gguf`	Q4_K_M	~16 GB	Imatrix-guided; good default on 24–32 GB GPUs

Recommendation: start with Q4_K_M or Q5_K_M. Use Q8_0 / F16 if you are re-quantizing or measuring ceiling quality.

What changed vs Fusion BF16

Heretic abliteration (v1.4.0) on the BF16 weights
- 200 optimization trials
- Scorers: KeywordRate + KLDivergence
- Eval sets: mlabonne/harmful_behaviors + mlabonne/harmless_alpaca
- Selected trial 77: refusal keywords 89 → 3 / 100, KL divergence 0.026
- Across trials: KL mean ≈ 0.008, max ≈ 0.073 (no trial > 0.1)
MTP / NextN preserved in the GGUF (blk.64 / nextn tensors bundled). Heretic’s merge-save dropped MTP safetensors; those tensors were grafted back from Fusion BF16 before convert.
Chat template: Fusion’s native template (not a Compat rewrite).
Quantization: llama.cpp convert → Q8_0 → imatrix (WikiText-2 calib, 100×256 chunks) → Q4_K_M and Q5_K_M with --imatrix.

Abliteration reduces refusal behavior; it does not remove your responsibility for how you use the model. Follow applicable law and the Qwen license.

Quick start

llama.cpp

bash

./llama-server \
  -m Qwopus3.6-27B-Fusion-heretic-MTP-Q4_K_M.gguf \
  -ngl 99 -c 32768 --jinja \
  --spec-type draft-mtp --spec-draft-n-max 4

LM Studio / koboldcpp / etc.

Import the .gguf of your choice. Prefer builds that understand Qwen3.6 / qwen35 and MTP speculative decoding if you want draft speedups.

OpenAI-compatible (example)

Point any OpenAI-compatible client at a llama-server / llama-swap endpoint serving one of these files. Thinking mode follows the Fusion / Qwen3.6 chat template (enable_thinking / jinja).

Model details


Parameters	~27B
Architecture	Qwen3.6 / qwen35 hybrid (linear + periodic full attention)
Context	Up to 262,144 native (practical limit = VRAM + KV cache)
Vision	Not included in these text GGUFs (BF16 Fusion retains vision tensors)
MTP	Included (NextN / draft head)
License	Qwen (see license link in YAML)

Provenance

Step	Source / tool
Base	KyleHessling1/Qwopus3.6-27B-Fusion-BF16
Abliteration	Heretic v1.4.0
Convert / quant / imatrix	llama.cpp
Uploader	AtticusG3

Limitations

Smoke-tested for generation on a small LiveCodeBench prompt sample after quant; not a full re-benchmark of Fusion’s published scores.
Some MTP (blk.64) tensors lacked imatrix entries and fell back to plain K-quants.
Uncensored / abliterated models can produce disallowed content more readily than the stock chat model — use with care.

Acknowledgements

Kyle Hessling — Fusion merge and BF16 release
Qwen team — Qwen3.6-27B
p-e-w / Heretic — abliteration tooling
llama.cpp contributors — GGUF convert, imatrix, quantize

Citation

bibtex

@misc{qwopus36-fusion-heretic-mtp-gguf,
  title        = {Qwopus3.6-27B-Fusion-heretic-MTP GGUF},
  author       = {AtticusG3},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/AtticusG3/Qwopus3.6-27B-Fusion-heretic-MTP-GGUF}},
  note         = {Heretic abliteration of KyleHessling1/Qwopus3.6-27B-Fusion-BF16; imatrix GGUF quants}
}