Qwen3.8-27B-Heretic · GSQ-RCO GGUFs

Non-uniform GGUF quantizations of RVN Qwen3.8-27B-Heretic-Abliterated-Uncensored (refusals 0-1/100, KL ~ 0.0085 vs base). Instead of one uniform quantization type for all tensors, each of the 851 tensors carries its own type, following the per-tensor RCO allocation published by IST-DASLab for their GSQ-RCO release of Qwen3.8-27B, quantized locally with llama.cpp and the official importance matrix imatrix-qwen3.8-27b.gguf. Files are standard GGUF and run unmodified in llama.cpp / Ollama / LM Studio.

Why this release exists

This family was made at the request of community member Heidagger44 in discussion #11:

Regarding IQ3_S there is new method: GSQ-RCO, if it's possible to combine your abliteration method with this quantization, we would have a true refusal free Qwen3.8-27B running fully on 16GB cards.

Thanks for the pointer - this is that combination, in four sizes.

Available files

Quality - measured, not assumed

Perplexity on wikitext-2 (llama-perplexity, ctx 4096, one RTX PRO 6000 Blackwell, full offload):

At 3.50 bpw the allocation lands within 1% of F16 - and a same-class uniform quantization sits 1.9% behind it at a larger file size (12.42 GB vs 11.80 GB). Generation speeds (same box, full offload): IQ2_XS 105 t/s, IQ2_S 101 t/s, IQ3_XXS 97 t/s, IQ3_S 90 t/s, IQ3_S-mtp 150 t/s with --spec-type draft-mtp.

How these files were built

1. Allocation - IST-DASLab's published per-tensor RCO maps for this tier set (851 tensor-to-type entries each; targets 2.50-3.50 bpw).
2. Quantization - every tensor of RVN-F16.gguf quantized at its assigned type with llama.cpp (--tensor-type-file), using the official imatrix-qwen3.8-27b.gguf.
3. MTP embedding - the official Qwen3.8 MTP/NextN head (from mtp-RVN.gguf) raw-byte merged into the -mtp twins (block_count=65, nextn_predict_layers=1).

Exactness - header audit per file: 803/851 tensors match the published allocation exactly; the 48 ssm_alpha tensors are held at F32 by llama.cpp's SSM precision rule (the map lists BF16 - strictly higher precision, ~ +24 MB/file).

Provenance note - this reproduces the RCO allocation with optimized local quantization. The per-tensor GSQ refinement pass from ISTA's internal pipeline is not reproduced; its current public implementation is a research checkpoint that regresses end-to-end KL in its own tests. The allocation and the official importance matrix are the published, verifiable ingredients. If the KL-gated refinement lands upstream we will re-run it.

Usage

# 16 GB cards: start here
llama-cli -m RVN-Qwen3.8-27B-Heretic-GSQ-RCO-IQ3_S.gguf -ngl 99 --temp 1.0 -p "Your prompt"

# speculative decoding with the embedded MTP head (llama.cpp >= b10440):
llama-cli -m RVN-Qwen3.8-27B-Heretic-GSQ-RCO-IQ3_S-mtp.gguf -ngl 99 \
  --spec-type draft-mtp --spec-draft-n-max 2 --parallel 1 -p "Your prompt"

Thinking mode follows the model's chat template. A vLLM/Transformers (compressed-tensors) GSQ variant lives in the companion repository under GSQ-3bit/.

Related repositories

• Main model + standard quant spectrum: Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF
• FP16 safetensors (vLLM/SGLang): Qwen3.8-27B-Heretic-Abliterated-Uncensored

Credits

Per-tensor RCO allocation + official importance matrix: IST-DASLab (GSQ-RCO release). Quantization, verification, and MTP embedding: this repository. Base abliteration (RVN / ARA): see the main repository.