license: apache-2.0
language:
- en
- zh
- multilingual
tags: - text-generation
- image-text-to-text
- gguf
- llama.cpp
- safetensors
- qwen
- qwen3
- qwen3.5
- qwen3.6
- claude-opus-distill
- reasoning
- vision
- multimodal
- abliterated
- refusal-ablated
- uncensored
- q8_0
- conversational
base_model: - Jackrong/Qwopus3.6-27B-v2
- Qwen/Qwen3.6-27B
pipeline_tag: image-text-to-text
library_name: gguf
Qwen3.6-27B-V2-abliterated-uncensored-8-bit-GGUF
Yes — MULTIMODAL. Bundled
mmproj.gguf(~928 MB, F16) preserves the full Qwen3.6-VL vision tower. Use it withllama-server --mmprojorllama-mtmd-clifor text + image inference.
Q8_0 (8-bit, 8.50 BPW) of a abliterated Qwen 3.6 27B v2 (the Jackrong Claude-Opus reasoning distill of Qwen 3.6 27B). Refusals reduced from 91/100 → 4/100 with KL drift of just 0.0176. By the Lemura Labs research team.
TL;DR
| Property | Value |
|---|---|
| Disk size | ~28 GB (27 GB LM + 928 MB mmproj) |
| BPW | 8.50 (Q8_0) |
| Scheme | llama.cpp Q8_0 — symmetric 8-bit per-block scale, no FFN downcasting. |
| Refusal rate (the ablation toolkit, n=100) | 4/100 (vs vanilla Qwen 3.6 91/100) |
| KL divergence vs vanilla (at BF16) | 0.0176 |
| Vision | Yes — via paired mmproj.gguf |
| Recommended RAM/VRAM | 36 GB+ Apple Silicon / 32 GB GPU |
| Runtime | stock ggml-org/llama.cpp (any recent build) — no custom fork needed for Q8_0. |
| Released by | Lemura Labs |
All Qwen3.6-27B variants
The full Qwen3.6-27B family from Lemura Labs — same abliterated weights (refusal 4/100, KL 0.0176), different quant schemes for different runtimes.
| Quant | Format | BPW | Disk | Vision | Runtime | Link |
|---|---|---|---|---|---|---|
| 8-bit | MLX | 8.50 | ~27 GB | Yes — native | mlx-vlm | …-8-bit-mlx |
| 6-bit | MLX | 6.66 | ~21 GB | Yes — native | mlx-vlm | …-6-bit-mlx |
| OptiQ 3.7bpw | MLX | ~3.7 | ~14 GB | Yes — ViT spliced | mlx-vlm | …-OptiQ-3.7bpw-mlx |
| Q8_0 (this repo) | GGUF | 8.50 | ~28 GB | Yes — via mmproj | llama.cpp | — (you are here) |
| Q6_K | GGUF | ~6.56 | ~22 GB | Yes — via mmproj | llama.cpp | …-6-bit-GGUF |
| Q4_K_M | GGUF | ~4.92 | ~16 GB | Yes — via mmproj | llama.cpp | …-Q4_K_M-GGUF |
| TQ3_4S | GGUF | 4.00 (~3.5 eff) | ~14 GB | Yes — via mmproj | llama.cpp-tq3 | …-TQ3_4s-GGUF |
| TQ3_1S | GGUF | 4.00 (~3.5 eff) | ~14 GB | Yes — via mmproj | llama.cpp-tq3 | …-TQ3_1s-GGUF |
All variants share the same abliterated base weights — pick by your runtime (Apple Silicon → MLX; CUDA/CPU/cross-platform → GGUF) and your RAM budget.
Lineage
Qwen/Qwen3.6-27B (Qwen Team — base multimodal pretrain)
│
▼
Jackrong/Qwopus3.6-27B-v2 (Jackrong — Claude-Opus reasoning distill)
│
▼
ablation abliteration (TPE-50) (Lemura Labs)
├── 25 random startup trials
├── 2 community priors (coder3101, wangzhang)
└── 23 TPE smart-sampling trials → best at trial 45
│
▼
HF safetensors → F16 GGUF via llama.cpp-tq3 (Lemura Labs)
│
▼
this repo — Qwen3.6-27B-V2-abliterated-uncensored-Q8_0 GGUF + paired mmproj.gguf
Direct upstream links:
- Foundation: Qwen/Qwen3.6-27B
- Claude-Opus distill: Jackrong/Qwopus3.6-27B-v2
- Abliteration tool: the ablation toolkit by Lemura Labs
- Quantization tool: turbo-tan/llama.cpp-tq3 (a fork of ggml-org/llama.cpp)
Abliteration Results
| Stage | Refusals (n=100) ↓ | KL divergence ↓ |
|---|---|---|
| Vanilla Jackrong/Qwopus3.6-27B-v2 | 91 / 100 | — (reference) |
| Community prior: coder3101 (T27) | 4 / 100 | 0.0359 |
| Community prior: wangzhang (T28) | 30 / 100 | 0.0259 |
| TPE best (T45) — shipped here | 4 / 100 | 0.0176 |
| TPE second-best (T37) | 5 / 100 | 0.0210 |
→ 96% reduction in refusals with capability preserved (KL ≈ 0.018, well below the 0.3 healing threshold). No SFT / LoRA healing was required.
Method (TPE-50 with community priors → llama.cpp GGUF)
Step 1. Abliteration (the ablation toolkit TPE-50, BF16 source)
- 25 random startup trials + 2 community priors enqueued (coder3101 dir=37.97, wangzhang dir=34.66) + 23 TPE smart-sampling trials.
- Best Pareto trial: T45 (
direction_index=41.42) — 4/100 refusals at KL=0.0176. - Auto-saved via the ablation toolkit's LoRA-adapter merge path with vision tower fully intact.
Total the ablation toolkit wall-clock: ~13 h on M4 Max 128 GB.
Step 2. HF safetensors → F16 GGUF
python convert_hf_to_gguf.py \
/path/to/Qwen 3.6 27B-v2-abliterated \
--outfile Qwen 3.6 27B-v2-abliterated-F16.gguf \
--outtype f16
The turbo-tan fork's converter registers Qwen3_5ForConditionalGeneration natively and emits proper SSM tensors (ssm_a, ssm_conv1d, ssm_alpha, ssm_beta, ssm_out) alongside the gated-attention layers.
Step 3. Vision tower → mmproj.gguf
python convert_hf_to_gguf.py \
/path/to/Qwen 3.6 27B-v2-abliterated \
--outfile mmproj-Qwen 3.6 27B-v2-abliterated-F16.gguf \
--outtype f16 \
--mmproj
This emits a separate 928 MB GGUF containing the 27-block Qwen3-VL ViT (334 vision tensors at F16/F32) plus the multimodal projector.
Step 4. Quantization
./build/bin/llama-quantize \
Qwen 3.6 27B-v2-abliterated-F16.gguf \
Qwen3.6-27B-V2-abliterated-uncensored-Q8_0.gguf \
Q8_0
Use it
llama-server (OpenAI-compatible HTTP, multimodal)
./build/bin/llama-server \
-m Qwen3.6-27B-V2-abliterated-uncensored-Q8_0.gguf \
--mmproj mmproj-Qwen 3.6 27B-v2-abliterated-F16.gguf \
--host 127.0.0.1 --port 8080 \
-ngl 99 -c 8192 -fa on --jinja
Then point any OpenAI-compatible client at http://127.0.0.1:8080/v1.
llama-mtmd-cli (one-shot multimodal generation)
./build/bin/llama-mtmd-cli \
-m Qwen3.6-27B-V2-abliterated-uncensored-Q8_0.gguf \
--mmproj mmproj-Qwen 3.6 27B-v2-abliterated-F16.gguf \
--image photo.jpg \
-p "Describe this image briefly."
llama-cli (text-only)
./build/bin/llama-cli \
-m Qwen3.6-27B-V2-abliterated-uncensored-Q8_0.gguf \
-ngl 99 \
-c 8192 \
--jinja \
-p "Explain the difference between SSM and softmax attention in three sentences."
Ollama / LM Studio / Jan
Drop the two GGUF files into the runtime's models directory; standard multimodal flow.
Quantization details
- Source weights: BF16 abliterated checkpoint (12 shards, ~50 GB) — the ablation toolkit T45 merged into
Jackrong/Qwopus3.6-27B-v2. - Intermediate: F16 GGUF (53.8 GB, 851 tensors) produced by
convert_hf_to_gguf.pyfromturbo-tan/llama.cpp-tq3. - Final quantization: see Step 4 above.
- Vision projector: F16, 928 MB, shipped as
mmproj-Qwen 3.6 27B-v2-abliterated-F16.ggufin this repo. Mandatory for image input; standard llama.cpp--mmprojflag.
Architecture notes
Qwen 3.6 27B uses a hybrid attention stack — 3 GatedDeltaNet (linear attention / SSM) layers followed by 1 full-softmax-attention layer, repeated 16× for 64 total layers; hidden 5120, vocab 248320, context 262144. The hybrid arch is supported in the turbo-tan/llama.cpp-tq3 fork (the upstream Qwen3_5ForConditionalGeneration registration). The SSM kernels run via llama.cpp's ssm_* tensor types.
Behavior caveats
- Uncensored. Refusal directions were surgically removed; this model will answer prompts the parent would refuse. Use responsibly and within applicable law. The release is provided for safety research, red-teaming, and creative/educational use cases.
- Multimodal preserved. Pair the LM GGUF with
mmproj.gguf(in this repo) to get full vision input. Without mmproj, the model still loads as text-only. - Identity preserved. The model still self-identifies as Qwen (developed by Alibaba's Tongyi Lab) — abliteration does not rewrite factual self-knowledge.
- Heavy chain-of-thought. Qwen 3.6 inherits Claude-Opus's verbose reasoning style. For terse answers, use a system prompt like
"Be brief and direct. Skip your reasoning.".
Credits
Quantization & release — Lemura Labs
Claude-Opus reasoning distill — Jackrong (Jackrong/Qwopus3.6-27B-v2)
Foundation model — Qwen Team @ Alibaba Tongyi Lab (Qwen/Qwen3.6-27B)
Abliteration toolkit — the ablation toolkit by Lemura Labs
Community priors — coder3101/Qwen3.5-27B-zerofuse · wangzhang/Qwen3.6-27B-abliterated
Runtime / converter — turbo-tan/llama.cpp-tq3 · ggml-org/llama.cpp
License
Apache-2.0, inherited from the foundation (Qwen3.6-27B) and the distill (Qwen 3.6 27B-v2) upstream.
Need a hosted endpoint, custom quant, or larger-scale inference? Lemura Labs — multi-provider LLM routing for the Indian developer ecosystem.