license: apache-2.0
base_model: 0xSero/Qwen3.5-99B
library_name: gguf
pipeline_tag: text-generation
tags:
- gguf
- abliterated
- qwen3.5
- moe
- reap
- not-for-all-audiences
Qwen3.5-122B-A10B-REAP-20 - Abliterated GGUF
Abliterated, importance-matrix (imatrix) quantized GGUFs of
0xSero/Qwen3.5-99B -
0xSero's ~20% MoE expert-prune (REAP) of Qwen/Qwen3.5-122B-A10B,
taking the model from 122B down to ~99B parameters (205 of 256 experts kept) while keeping the A10B
active-expert budget and the qwen35moe architecture. Robinson Labs then abliterated the pruned model
and quantized it here.
This is the lighter-prune sibling of the REAP-30 build:
20% of experts removed instead of 30%, so it keeps more of the original model's capacity at a larger
footprint. Like the REAP-30 variant it has no Multi-Token Prediction (MTP / NextN): the upstream
config declared a phantom nextn layer carrying no weights, so our convert produced a clean 48-layer
model (block_count=48). These are standard single-token-prediction GGUFs.
A full-precision bf16 safetensors base for re-abliteration, LoRA merge, fine-tune, or rolling your own
quants is at RobinsonLabs/Qwen3.5-122B-A10B-REAP-20-abliterated.
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.
Capability is preserved. Tagged not-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 |
|---|---|---|---|
Qwen3.5-122B-A10B-REAP-20-abliterated-Q6_K.gguf |
Q6_K | 81.3 GB | near-lossless |
Qwen3.5-122B-A10B-REAP-20-abliterated-Q5_K_M.gguf |
Q5_K_M | 70.4 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-Q4_K_M.gguf |
Q4_K_M | 60.2 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-Q4_K_S.gguf |
Q4_K_S | 56.5 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-IQ4_XS.gguf |
IQ4_XS | 53.2 GB | quality/size sweet spot |
Qwen3.5-122B-A10B-REAP-20-abliterated-Q3_K_M.gguf |
Q3_K_M | 47.6 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-IQ3_M.gguf |
IQ3_M | 43.7 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-IQ3_XS.gguf |
IQ3_XS | 40.9 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-IQ2_M.gguf |
IQ2_M | 32.8 GB | |
Qwen3.5-122B-A10B-REAP-20-abliterated-IQ2_XS.gguf |
IQ2_XS | 29.5 GB | smallest |
All quants are imatrix-weighted (generic calibration). This is the REAP-pruned 48-layer model - no
MTP block.
Method
- Expert prune (REAP): ~20% of the MoE experts removed by 0xSero's REAP method, 122B -> ~99B
params (205 of 256 experts),qwen35moearch, A10B active budget retained. - Abliteration: single mid-layer refusal direction removed via weight orthogonalization on the
bf16 pruned base; routers preserved. No MTP/NextN block exists in this variant. - Quant: importance-matrix (imatrix) weighted convert + quantize with
llama.cpp. The imatrix uses a generic calibration set
(bartowskicalibration_datav3) - a broad, domain-agnostic fit.
bf16 base
The full-precision bf16 safetensors base for this ladder is
RobinsonLabs/Qwen3.5-122B-A10B-REAP-20-abliterated
- the master for further surgery (re-abliteration, LoRA merge, fine-tune) and for making your own
quants. The upstream REAP parent is 0xSero/Qwen3.5-99B.
Provenance
Qwen3.5-122B-A10B (Apache-2.0) -> REAP-20 expert-prune (0xSero) -> abliterated bf16 (Robinson Labs)
-> generic-imatrix GGUF quants.