license: apache-2.0
base_model: OpenMOSE/Qwen3.5-REAP-262B-A17B
library_name: gguf
pipeline_tag: text-generation
tags:
- gguf
- abliterated
- uncensored
- imatrix
- moe
- reap
- qwen3.5
- not-for-all-audiences
Qwen3.5-REAP-262B-A17B Abliterated v2 GGUF
Status: building. This is the corrected re-abliteration of
Qwen3.5-REAP-262B-A17B-abliterated-GGUF
(v1). Rungs upload here as they finish, IQ3_XS first.
Why v2 exists
A community report on v1 showed the abliteration was weaker than intended: some prompts that
should pass were still refused, especially with thinking enabled. We reproduced it and confirmed
the uploaded files were the correct ones, so it was not a bad upload. The root cause was the
refusal direction. v1 reused the direction captured on the sibling 212B model instead of capturing
one on this 262B directly, so the reduction was marginal.
What changed in v2
- The refusal direction is captured natively on the 262B itself, not borrowed from the 212B.
- An attention-sink guard rejects any direction dominated by a massive-activation channel, so the
orthogonalization removes the refusal reflex without damaging coherence. - Everything else matches the v1 method: single-direction weight orthogonalization of the residual
writers, harm guardrails retained by design, imatrix-weighted IQ rungs cut from a Q8_0 master.
Provenance
Qwen3.5-397B-A17B (Apache-2.0) -> OpenMOSE/Qwen3.5-REAP-262B-A17B (34% REAP prune) ->
abliterated v2 (native direction, sink-guarded) -> Q8_0 master -> quant rungs.
Base weights and full method: see v1's card and the base repo
RobinsonLabs/Qwen3.5-REAP-262B-A17B-abliterated.
Built by Robinson Labs.