license: apache-2.0
base_model: Qwen/Qwen3.6-35B-A3B
library_name: gguf
pipeline_tag: text-generation
tags:
- gguf
- abliterated
- qwen3.6
- moe
- mtp
- not-for-all-audiences
Qwen3.6-35B-A3B - Abliterated GGUF
Abliterated, importance-matrix (imatrix) quantized GGUFs of
Qwen/Qwen3.6-35B-A3B - a 35B-parameterqwen35moe MoE with an A3B active-expert budget. Robinson Labs abliterated the base model and
quantized it here.
Multi-Token Prediction (MTP / NextN) is preserved through abliteration, conversion, and
quantization: the blk.40.nextn.* tensors are intact (41-block model), so the speculative-decode
path is available to runtimes that support it.
Disclosure
This model is abliterated - the hard-refusal reflex on adult / creative content has been
reduced via single-direction weight orthogonalization. Harm guardrails are retained by design:
self-harm prompts still redirect to help (e.g. 988), and it is not intended to assist genuine
wrongdoing. This is a v1, partial abliteration; capability is preserved. Taggednot-for-all-audiences. Use responsibly - you are responsible for your use. License inherited from
the base model: Apache-2.0.
Files
| File | Quant | ~Size | Notes |
|---|---|---|---|
Qwen3.6-35B-A3B-abliterated-Q6_K.gguf |
Q6_K | ~27 GB | near-lossless |
Qwen3.6-35B-A3B-abliterated-Q5_K_M.gguf |
Q5_K_M | ~24 GB | |
Qwen3.6-35B-A3B-abliterated-Q5_K_S.gguf |
Q5_K_S | ~23 GB | |
Qwen3.6-35B-A3B-abliterated-Q4_K_M.gguf |
Q4_K_M | ~20 GB | |
Qwen3.6-35B-A3B-abliterated-Q4_K_S.gguf |
Q4_K_S | ~19 GB | |
Qwen3.6-35B-A3B-abliterated-IQ4_XS.gguf |
IQ4_XS | ~18 GB | quality/size sweet spot |
Qwen3.6-35B-A3B-abliterated-Q3_K_M.gguf |
Q3_K_M | ~16 GB | |
Qwen3.6-35B-A3B-abliterated-IQ3_M.gguf |
IQ3_M | ~15 GB | |
Qwen3.6-35B-A3B-abliterated-Q3_K_S.gguf |
Q3_K_S | ~15 GB | |
Qwen3.6-35B-A3B-abliterated-IQ3_XS.gguf |
IQ3_XS | ~14 GB | |
Qwen3.6-35B-A3B-abliterated-IQ2_M.gguf |
IQ2_M | ~12 GB | |
Qwen3.6-35B-A3B-abliterated-IQ2_XS.gguf |
IQ2_XS | ~10 GB | smallest |
All quants are MTP-preserved and imatrix-weighted (generic calibration).
Q5_K_S and Q3_K_S were added 2026-08 to complete the ladder. They are cut from the same Q8_0
master and the same calibration set as the other ten rungs, using a faithfully reproduced
imatrix (identical model, calibration data, and chunk count; the original imatrix file was
lost before it was archived). Their weighting basis is near-identical, not bit-identical, to
the original rungs.
Method
- Abliteration: single mid-layer refusal direction removed via weight orthogonalization on the
bf16 base; MTP (nextn) block and routers preserved. - Quant: importance-matrix (imatrix) weighted convert + quantize with
llama.cpp. The imatrix here uses a generic calibration
set (bartowskicalibration_datav3) - a broad, domain-agnostic fit. The MTP/NextN layer
(blk.40) is pinned to Q6_K on the very-low-bit rungs (it has no imatrix coverage).
bf16 base
The full-precision bf16 safetensors base for this ladder is
RobinsonLabs/Qwen3.6-35B-A3B-abliterated
- the master for further surgery (re-abliteration, LoRA merge, fine-tune) and for making your own
quants. The upstream base is Qwen/Qwen3.6-35B-A3B.
Provenance
Qwen3.6-35B-A3B (Apache-2.0) -> abliterated bf16 (Robinson Labs) -> Q8_0 master ->
generic-imatrix GGUF quants.