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 on the bf16 base: the rank-1 refusal
component is subtracted from every residual-stream writer (
o_proj, DeltaNet out_proj, fused +
shared-expert down_proj, token embedding). 181 tensors edited; routers and norms untouched. The
refusal direction is massive-activation guarded — attention-sink dimensions that would brick
the model are detected and excluded, and the direction is drawn from the clean late-layer
consensus (KB #213).
- Format — safetensors, sharded (122 shards), with config + tokenizer + index.
Architecture
qwen3_5_moe hybrid: 60 decoder layers (45 linear-attn / DeltaNet + 15 full-attn,
full_attention_interval=4), 267 experts, hidden size 4096, ~211.8B params / ~17B active. Text path
only. The REAP prune drops the native MTP head; the GGUF ladder corrects the phantom MTP layer at
convert time so it loads as a clean 60-layer decoder.
Files
Table with columns: Format, Precision, ~Size, Notes| Format | Precision | ~Size | Notes |
|---|
| safetensors (sharded, 122 shards) | bf16 | ~424 GB | abliterated master |
Quants
GGUF quants (Q5_K_M down to IQ2_XXS, plus a cluster-fit attention-priority mix, all
imatrix-weighted) are published at
RobinsonLabs/Qwen3.5-REAP-212B-A17B-abliterated-GGUF.
Provenance
Qwen3.5-397B-A17B (Apache-2.0) → OpenMOSE/Qwen3.5-REAP-212B-A17B (48% REAP prune) →
abliterated (bf16) → GGUF quant ladder.