license: apache-2.0 base_model: OpenMOSE/Qwen3.5-REAP-262B-A17B library_name: transformers pipeline_tag: text-generation tags:

• abliterated
• uncensored
• reap
• qwen3.5
• moe
• not-for-all-audiences

Qwen3.5-REAP-262B-A17B - Abliterated (bf16 base)

Abliterated bf16 base of OpenMOSE/Qwen3.5-REAP-262B-A17B, in safetensors, itself a 34% REAP expert-pruning of Qwen3.5-397B-A17B down to 262B total / ~17B active.

This is the full-precision master behind the RobinsonLabs/Qwen3.5-REAP-262B-A17B-abliterated-GGUF quant ladder (the ladder itself is cut from a Q8_0 quant master converted from this base). If you want a ready-to-run quant, use that repo. This repo is the master for further surgery (re-abliteration, LoRA merge, fine-tune) and for making your own quants.

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 what you generate with it. License inherited from the base model: Apache-2.0.

Method

• Abliteration: single-direction weight orthogonalization (FailSpy / Labonne method) on the bf16 base. For every matrix that writes the residual stream (o_proj, DeltaNet out_proj, fused expert down_proj, shared-expert down_proj, and the token embedding), the rank-1 component along the refusal direction is subtracted. Routers and norms pass through byte-identical.
• Refusal direction, massive-activation guarded. The direction is captured with a mean-difference control vector, then guarded against attention-sink contamination: the sink dimensions that dominate raw activation magnitude (and would brick the model if ablated) are detected across layers and excluded, and the direction is taken from the clean, spread-out consensus of the late layers rather than a single sink-dominated layer.
• Format: safetensors, sharded (61 shards), with config + tokenizer + index.

Architecture

qwen3_5_moe hybrid: 60 decoder layers (45 linear-attn / DeltaNet + 15 full-attn, full_attention_interval=4), 333 experts with 10 active per token, hidden size 4096, head dim 256, ~261.6B params / ~17B active, 262144 native context. No MTP / NextN layer. This is the text path only (no vision mmproj).

The high expert count is the defining feature of this REAP tier: 333 experts versus 267 on the 48%-pruned 212B sibling. More experts retained means more of the 397B parent's routing diversity survives, at the cost of size.

Files

Quants

GGUF quants (Q5_K_M down to IQ2_XXS, the IQ rungs imatrix-weighted on a domain corpus) are published at RobinsonLabs/Qwen3.5-REAP-262B-A17B-abliterated-GGUF.

Provenance

Qwen3.5-397B-A17B (Apache-2.0) -> OpenMOSE/Qwen3.5-REAP-262B-A17B (34% REAP prune) -> abliterated (bf16 master) -> Q8_0 master -> GGUF quant ladder.

Built by Robinson Labs.