Abliterated bf16 safetensors base 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 in-house. This repo is the full-precision master, in safetensors.

This is the lighter-prune sibling of the REAP-30 base: 20% of experts removed instead of 30%, keeping more of the original capacity at a larger footprint.

This is the bf16 base that the RobinsonLabs/Qwen3.5-122B-A10B-REAP-20-abliterated-GGUF quant ladder was quantized from. 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 rolling your own quants.

Like the REAP-30 variant, this build has no Multi-Token Prediction (MTP / NextN). The real model is a clean 48-layer qwen35moe - standard single-token prediction.

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.

Method

Expert prune (REAP): ~20% of the MoE experts removed by 0xSero's REAP method, 122B -> ~99B params (205 of 256 experts), qwen35moe arch, A10B active budget retained. Done upstream, in 0xSero/Qwen3.5-99B.
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.
Format: safetensors, sharded, with config + tokenizer + index. No vision tensors, no MTP head.

Files

Format	Precision	~Size	Notes
safetensors (50 shards)	bf16	~185 GB	abliterated REAP base; `qwen35moe`, 48 layers, no `nextn`

The model is 48 transformer layers (block_count=48), qwen35moe architecture, ~99B params with an A10B active-expert budget. The upstream config declares a phantom nextn (MTP) layer that carries no weights - downstream GGUF converters should treat this as a clean 48-layer model (block_count=48) and ignore the phantom head.

Quants

GGUF quants (Q6_K down to IQ2_XS, imatrix-weighted, 48-layer / no MTP) are published at RobinsonLabs/Qwen3.5-122B-A10B-REAP-20-abliterated-GGUF.

Provenance

Qwen3.5-122B-A10B (Apache-2.0) -> REAP-20 expert-prune (0xSero, repo "Qwen3.5-99B") -> abliterated bf16 (Robinson Labs). This safetensors repo is the abliterated bf16 master; the GGUF ladder is quantized from it.