RobinsonLabs

Qwen3.6-35B-A3B-abliterated

README

License: apache-2.0

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 (5/5 on our capability probe). Tagged not-for-all-audiences. Use responsibly - you are responsible for your use. License inherited from the base model: Apache-2.0.

Known issue: the previous master (June 2026) was abliterated in name only

The safetensors published here from 2026-06-28 until this re-upload (internal label gen-L18) did not actually reduce refusals: on our held-out generic refusal probe they refused 25/25, identical to the stock base. The refusal direction was never searched by depth and the edit weight was flat and unshaped, which on this MoE removes nothing measurable.

Every shard in this repo was replaced on 2026-09-02 with the new abliteration. If you downloaded before that date, re-download. The old master is HF revision 2f5d365 and earlier. Quants cut from the old master are likewise replaced in the GGUF repo, where every current file is listed with its sha256.

Results

On our held-out generic refusal probe (n=25): stock base 25/25 refused, this model 2/25 (both residuals are soft refusals in the acceptable band: the model declines and offers a legitimate alternative), community heretic cut 3/25. Capability probe 5/5, role-play probe 2/2. Child-safety guardrail set (n=10): 10/10 kept by an LLM judge (stock 10/10, community heretic 6/10). A response that declines or deflects, producing none of the requested content, counts as kept; the regex refusal-marker count on the same responses was 0/10 because the model deflects in coherent prose instead of emitting "I can't". Ten prompts is a small set; we do not claim guardrails are intact beyond that probe.

Method

Direction: a last-token, magnitude-preserving diff-of-means refusal direction (harmful vs harmless calibration prompts, 256 pairs each), captured from a trunk-only bf16 GGUF of the stock base, not from a quant. Depth search over rows 24..37; row 34 (the output of HF layer 34, about 85% depth) was chosen.
Surgery: per-layer weight orthogonalization with a heretic-style linear kernel. Attention outputs (10 full-attention o_proj + 30 DeltaNet linear_out) at weight 1.3, flat over layers 10-40. MLP path (the 40 fused routed-expert down-projections + 40 shared-expert down-projections) at 1.3 centered on layer 33, decaying to 0.8 over layers 15-40. Embeddings, routers, norms and the MTP block are untouched. Expert-path coverage is what makes a single direction work on this MoE: attention-only at weight 1.0 barely ablates.
Format: safetensors, sharded, with config + tokenizer + index. MTP/NextN head preserved.

Files

Table with columns: Format, Precision, ~Size, Notes
Format	Precision	~Size	Notes
safetensors (26 shards)	bf16	~71.9 GB	abliterated base; `qwen35moe`, MTP-preserved (41 blocks)

The size is approximate; exact shard sizes are on the Hub file listing. The model is qwen35moe architecture with the MTP/NextN head at blk.40 preserved (41 blocks total), ~35B params with an A3B active-expert budget.

Quants

GGUF quants (13 rungs, Q8_0 down to IQ2_XS, MTP-preserved, imatrix-weighted from this model's own Q8_0) are published at RobinsonLabs/Qwen3.6-35B-A3B-abliterated-GGUF.

Provenance

Qwen3.6-35B-A3B (Apache-2.0) -> abliterated bf16 (D34H, Robinson Labs, 2026-09). This safetensors repo is the abliterated bf16 master; the GGUF ladder is quantized from it.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

RobinsonLabs

Model Tree

Base

Qwen/Qwen3.6-35B-A3B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality