license: apache-2.0
base_model:
- llmfan46/Qwen3.6-35B-A3B-uncensored-heretic
base_model_relation: quantized
library_name: gguf
pipeline_tag: text-generation
tags: - gguf
- macaq
- macaq-r
- qwen3_5_moe
- mixture-of-experts
- heretic
- uncensored
- abliterated
Qwen3.6-35B-A3B-uncensored-heretic — Q5_K_M MACaQ-R (GGUF)
A MACaQ-R refined version of the Q5_K_M quant of
llmfan46/Qwen3.6-35B-A3B-uncensored-heretic.
MACaQ-R (Huawei CSL, paper) does not
quantize from scratch: it refines the weight codes of an existing native K-quant MoE GGUF (Q4_K / Q5_K / Q6_K) against a
full-precision teacher, keeping block scales, tensor formats and the original activation basis. All other tensors are
copied unchanged. The result loads in standard llama.cpp.
Both the HF reference checkpoint (--hf) and the teacher (--teacher) were built from the heretic weights, not from
stock Qwen3.6, so the refinement targets the abliterated model's own distribution.
File
| File | Size | SHA-256 |
|---|---|---|
Qwen3.6-35B-A3B-uncensored-heretic-Q5_K_M-MACaQ-R.gguf |
24,764,032,768 bytes | d4f9aabd08f8cec3580c3d53d60b28a450e86d6f7c5404014371eff2a3d2a15c |
How it was made
- Source quant:
Qwen3.6-35B-A3B-uncensored-heretic-Q5_K_M.gguffrom
llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF - HF reference:
llmfan46/Qwen3.6-35B-A3B-uncensored-heretic(bf16 safetensors) - Teacher: f16 GGUF converted from the same heretic safetensors (
convert_hf_to_gguf.py --outtype f16 --no-mtp) - Calibration:
examples/make_moe_calibration.py --profile diverse --seed 617029 --count 6000 - Refine command (MACaQ
llamacpp/macaq.py, defaults otherwise: context 512, 256 chunks, 64 iterations):
python llamacpp/macaq.py refine \
--model Qwen3.6-35B-A3B-uncensored-heretic-Q5_K_M.gguf \
--hf models/heretic-hf --teacher models/heretic-f16.gguf \
--calibration data/calibration.txt --gpus 0,1 --device cuda \
--gpu-layers 28 --ubatch 512 --threads 48 \
--output Qwen3.6-35B-A3B-uncensored-heretic-Q5_K_M-MACaQ-R.gguf --work-dir work/refine
- Refined: 370 K-quant weight tensors (120 expert tensors × 256 experts).
- Hardware: 2× RTX 3090.
Usage
llama-server -m Qwen3.6-35B-A3B-uncensored-heretic-Q5_K_M-MACaQ-R.gguf -ngl 99
Smoke test (llama.cpp, 2× RTX 3090, -ngl 99): loads in ~18 s, ~139 tokens/s generation.
Evaluation status
Quality has not been measured yet for this file (no KL-divergence / perplexity numbers versus the f16 teacher, and
no check of refusal behaviour). For reference, Huawei reports for stock Qwen3.6-35B-A3B at the same 22.36 GB size a
KL-divergence to FP16 of 0.013367 for MACaQ-R Q4_K_XL vs 0.015731 for Unsloth UD-Q4_K_XL; that number is not a claim
about this file.
License
Apache-2.0, following the licenses of
Qwen/Qwen3.6-35B-A3B,
llmfan46/Qwen3.6-35B-A3B-uncensored-heretic
and MACaQ.
This is an uncensored (abliterated) model. You are responsible for how you use it.