license: apache-2.0
base_model:

• Qwen/Qwen3.6-27B
  tags:
• Qwen3.6-27B
• abliterated
• Uncensored
• Smaller
• TurboQuant
• gguf
• IQ4_XS
  pipeline_tag: image-text-to-text

Qwen3.6-27B-abliterated-i1-IQ4_XS-GGUF(Smaller)

** ♥♥♥♥ This model is specially prepared for your graphics card with 16GB VRAM. **

This model is equivalent to https://huggingface.co/lemonyins/Qwen3.6-27B-uncensored-abliterated-i1-IQ4_XS-GGUF-Smaller

Innovation

This model refers to the fully optimized Qwen3.6-27B-i1-IQ4_XS, which has restored the attn_qkv layer to pure IQ4_XS. Furthermore, we introduce a novel hybrid precision quantization strategy: the FFN layer uses IQ3_S, which achieves significantly smaller file sizes while maintaining core inference capabilities through support from TurboQuant KV caching. Additionally, we use an Huihui abliterated version of the base model for quantization, making it convenient for users to conduct in-depth research.

Motivation

The original llama.cpp quantization heuristics upgrade attn_qkv layers to q5_K under certain conditions (e.g., n_gqa >= 4), causing noticeable file bloat. cHunter789's fix restores attn_qkv to pure IQ4_XS, saving ~375 MiB.

Taking this further: FFN layers (ffn_down, ffn_up, ffn_gate) account for ~2/3 of total model parameters, yet they have higher redundancy than attention layers. Downgrading them from IQ4_XS to IQ3_S is a natural next step — it yields substantial size reduction with minimal quality impact, especially when attention layers (which dominate inference quality) remain at IQ4_XS.

Methodology

1. Base model: Huihui-Qwen3.6-27B-abliterated by mradermacher (abliterated, F16) — an uncensored version optimized for inference
2. Quantization tool: llama.cpp with TurboQuant support, built from TheTom/llama-cpp-turboquant
3. Quantization types:
   • attn_qkv, attn_k, attn_v, attn_output, output: IQ4_XS
   • ffn_down, ffn_up, ffn_gate: IQ3_S
   • Other layers: default IQ4_XS
4. Importance matrix (imatrix): sourced from mradermacher's Qwen3.6-27B-i1-GGUF

Quantization Commands

Option A — Pure IQ4_XS baseline:

llama-quantize.exe \
  --imatrix Huihui-Qwen3.6-27B-abliterated.imatrix.gguf \
  Huihui-Qwen3.6-27B-abliterated-BF16.gguf \
  Huihui-Qwen3.6-27B-abliterated-i1-IQ4_XS.gguf \
  IQ4_XS

Option B — Recommended: IQ4_XS-FFN-IQ3_S (smaller + longer context):

llama-quantize.exe ^
  --imatrix Huihui-Qwen3.6-27B-abliterated.imatrix.gguf ^
  Huihui-Qwen3.6-27B-abliterated-BF16.gguf ^
  Huihui-Qwen3.6-27B-abliterated-i1-IQ4_XS-FFN-IQ3_S.gguf ^
  IQ4_XS ^
  --tensor-type "blk.*.ffn_down" iq3_s ^
  --tensor-type "blk.*.ffn_up" iq3_s ^
  --tensor-type "blk.*.ffn_gate" iq3_s

Note on TurboQuant: This model is recommended to be used with llama.cpp (https://github.com/TheTom/llama-cpp-turboquant) that supports TurboQuant KV caching. TurboQuant allows the KV cache to use a separate, more compact quantization format (turbo4 / turbo3), dramatically reducing memory usage even when the model weights themselves remain at IQ4_XS. Of course, it is also possible to use vllm or other inference frameworks that support TurboQuant technology, but the author used llama.cpp for the test.

Memory Performance (with TurboQuant KV Cache)

Key Takeaways

• IQ4_XS baseline reaches 60-80K context before VRAM saturation
• IQ4_XS-FFN-IQ3_S extends context window to 160-200K — more than 2x — with the same VRAM budget, thanks to the smaller FFN weight footprint enabling more KV cache capacity
• Quality-critical attention layers remain at IQ4_XS, so the model's reasoning and instruction-following capabilities are largely preserved

Inference Speed

Tested on NVIDIA RTX 4060 Ti 16GB:

Vision Support (Optional)

This model supports vision input when paired with the Vision Modality Projector (mmproj-BF16.gguf). If you load the vision module, the available context window will be reduced by approximately 10K to accommodate the vision encoder's memory footprint.

Example with vision support:

llama-server.exe ^
  -m Huihui-Qwen3.6-27B-abliterated-i1-IQ4_XS-FFN-IQ3_S.gguf ^
  -mmproj mmproj-BF16.gguf ^
  -c 153840 ^
  -ngl 99 ^
  --flash-attn on ^
  --cache-type-k turbo4 ^
  --cache-type-v turbo4 ^
  --host 0.0.0.0

Caveats

• TurboQuant is mandatory: This model relies on TurboQuant KV cache for the listed memory figures. Standard llama.cpp builds without TurboQuant will consume significantly more VRAM.
• FFN layers at IQ3_S: While the quality impact is expected to be minimal for most tasks, some degradation may be observable in tasks that heavily depend on FFN-related capabilities (e.g., certain factual recall scenarios). The attention layers remain at IQ4_XS to preserve core inference quality.
• Verification pending: Perplexity benchmarks with the standard IQ4_XS baseline are planned to quantify the quality difference precisely.

🧠 Intelligence (Perplexity) Comparison

Test using Chinese novel:

Test with code:

How to Use

You need a TurboQuant-enabled llama.cpp build from TheTom/llama-cpp-turboquant.

Note: This model is quantized from an abliterated (uncensored) base model. The base model removes content restrictions for research and development purposes.

Recommended: IQ4_XS-FFN-IQ3_S version — 160K context on ~15 GB VRAM

llama-server.exe ^
  -m Huihui-Qwen3.6-27B-abliterated-i1-IQ4_XS-FFN-IQ3_S.gguf ^
  -c 163840 ^
  -ngl 99 ^
  --flash-attn on ^
  --cache-type-k turbo4 ^
  --cache-type-v turbo4 ^
  --host 0.0.0.0

Alternative: Pure IQ4_XS version — shorter context, slightly larger file

llama-server.exe ^
  -m Huihui-Qwen3.6-27B-abliterated-i1-IQ4_XS.gguf ^
  -c 65536 ^
  -ngl 99 ^
  --flash-attn on ^
  --cache-type-k turbo4 ^
  --cache-type-v turbo4 ^
  --host 0.0.0.0

Acknowledgments

• cHunter789 — for the attn_qkv → IQ4_XS fix and the original fully optimized IQ4_XS GGUF
• mradermacher — for the base abliterated model and imatrix
• TheTom — for llama-cpp-turboquant, the TurboQuant KV cache implementation
• llama.cpp — the ggml/llama.cpp team for the base quantization framework