gaston-parravicini/Qwen3.6-27B-Abliterated-MTP-GGUF

Qwen3.6-27B Abliterated + MTP GGUF

The first publicly available Qwen3.6-27B uncensored GGUF with native MTP speculative decoding.
Refusal-free at the weight level · Full MTP block grafted · ~70 t/s on RTX 3090 · No custom fork required

Published by Gastón Parravicini

---

Why this exists

When Qwen3.6-27B dropped, two things were true at the same time:

• The only abliterated versions available had no MTP support — the draft heads were stripped during the merge, killing speculative decoding speed
• The only MTP-enabled GGUFs were fully censored — original refusal behavior intact

Nobody had combined both. Doing it required writing custom patches to handle Qwen3.6's MTP tensor naming conventions and avoid GGUF metadata corruption. This release is the result of that work.

This was the first. It still has the most complete quant coverage.

---

What this release adds

Refusal suppression

Removed at the weight level using two-pass orthogonal-projection abliteration (abliterix + Optuna TPE). KL divergence of 0.024 vs the base model — well below the 0.05 threshold where quality degradation becomes measurable. General intelligence, reasoning, and tool use are fully intact.

Full MTP speculative decoding

The complete blk.64 MTP block — all 15 tensors — grafted from the official Unsloth release. MTP is baked into Qwen3.6-27B at training time, so the draft heads are native, not bolted on.

Config	Speed	Notes
No MTP	~45 t/s	baseline
MTP --spec-draft-n-max 3	~70 t/s	recommended
Draft acceptance rate	90–100%	context-dependent
Max tested context	150k tokens	RTX 3090, q4_0 KV

~55% speedup. No quality loss. No custom fork required — MTP support is now in llama.cpp mainline.

---

Quants

File	Size	VRAM	Notes
Q2_K	11 GB	12 GB	Fast, lower quality
Q3_K_M	13 GB	16 GB	Decent for 16GB GPUs
Q4_K_M	16 GB	24 GB	Recommended — best quality/speed tradeoff
Q5_K_M	19 GB	24 GB+	Noticeably better than Q4
Q6_K	21 GB	32 GB	High fidelity
Q8_0	28 GB	48 GB	Near-lossless

Mac users: memory is unified — M2 Max 32GB fits Q4_K_M comfortably.

All quants include the full MTP draft head block. Speculative decoding works across all tiers.

---

Requirements

MTP support is now in llama.cpp mainline. No custom fork needed.

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target llama-server llama-cli

For Mac (Metal), replace -DGGML_CUDA=ON with -DGGML_METAL=ON.

---

Usage

Standard server (OpenAI-compatible API)

./llama-server \
  -m Qwen3.6-27B-abliterated-Gaston-MTP-Q4_K_M.gguf \
  -ngl 99 \
  --flash-attn on \
  --ctx-size 32768 \
  --spec-type draft-mtp \
  --spec-draft-p-min 0.75 \
  --spec-draft-n-max 3 \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --host 0.0.0.0 \
  --port 8080 \
  --jinja \
  -n 16384

Agentic and coding workloads (thinking visible)

./llama-server \
  -m Qwen3.6-27B-abliterated-Gaston-MTP-Q4_K_M.gguf \
  -ngl 99 \
  --flash-attn on \
  --ctx-size 80000 \
  -b 2048 -ub 1024 \
  --temp 0.6 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.0 \
  --spec-type draft-mtp \
  --spec-draft-p-min 0.75 \
  --spec-draft-n-max 3 \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --host 0.0.0.0 \
  --port 8080 \
  --jinja \
  -n 16384 \
  --chat-template-kwargs '{"preserve_thinking":true}'

Recommended sampling

temperature: 0.6
top_p: 0.95
top_k: 20
min_p: 0.0

---

Validated use cases

This model has been tested in production with:

• PentAGI — autonomous multi-agent penetration testing, full tool call loop
• Qwen Code — agentic coding with MCP servers (Context7, SearXNG)
• OpenClaw — local agent framework with OpenAI-compatible API
• Standard OpenAI function calling via --jinja chat template

The abliteration does not affect tool call behavior, function calling, or structured output generation.

---

Abliteration quality

Refusal suppression was done with orthogonal projection + Optuna TPE optimization. The result is one of the cleanest abliterations publicly available for this model:

Metric	This model	Typical abliteration
KL divergence vs base	0.024	0.04–0.10
Refusals (out of 100)	~2/100	varies
Quality threshold	< 0.05 ✅	often exceeded

Lower KL divergence = closer to the original model's output distribution = less quality loss. At 0.024, the abliteration is essentially invisible to downstream tasks.

---

Base model benchmarks

Capabilities are fully preserved. Abliteration removes refusal directions only — not reasoning, coding, or general intelligence.

Coding & Agentic

Benchmark	Qwen3.5-27B	Qwen3.6-27B
SWE-bench Verified	75.0	77.2
SWE-bench Pro	51.2	53.5
SWE-bench Multilingual	69.3	71.3
Terminal-Bench 2.0	41.6	59.3
SkillsBench Avg5	27.2	48.2
LiveCodeBench v6	80.7	83.9

Knowledge & Reasoning

Benchmark	Qwen3.5-27B	Qwen3.6-27B
MMLU-Pro	86.1	86.2
GPQA Diamond	85.5	87.8
AIME 2026	92.6	94.1
HMMT Feb 26	84.3	84.3
SuperGPQA	65.6	66.0

Full benchmark details: official Qwen3.6-27B model card.

---

How this was built

Standard abliteration pipelines strip MTP tensors during the merge — there was no existing tool that preserved them. This required a custom multi-step process:

1. Download wangzhang/Qwen3.6-27B-abliterated BF16 safetensors — abliteration via two-pass orthogonal projection with Optuna TPE optimization (KL divergence 0.024)
2. Convert to GGUF Q8_0 using the llama.cpp converter
3. Graft the full blk.64 MTP block (15 tensors) from unsloth/Qwen3.6-27B-MTP-GGUF using a patched version of havenoammo's graft script
4. Three custom patches written to handle Qwen3.6's MTP tensor naming conventions and prevent duplicate KV metadata errors in the output GGUF
5. Quantize all tiers from the grafted Q8_0 using llama-quantize
6. Validate MTP activation in llama.cpp — confirmed draft acceptance rates above 90% across all context lengths tested

The three patches were the non-trivial part. Qwen3.6's tensor naming diverges from prior Qwen versions in ways that broke existing graft tooling. No existing public script handled this correctly at the time of release.

---

Credits

Component	Author
Base model	Alibaba Qwen Team
Abliteration weights	wangzhang / abliterix
MTP draft heads	Unsloth
Graft script (base)	havenoammo
MTP llama.cpp branch (original validation)	am17an
MTP llama.cpp mainline merge	llama.cpp contributors
Custom patches, integration, quantization, publication	Gastón Parravicini

---

License

Apache 2.0, inherited from Qwen/Qwen3.6-27B.

---

Disclaimer

This model has had its safety guardrails removed at the weight level. It is intended for research, security testing, and unrestricted local inference. You are solely responsible for how you use it.