Qwen3.8-27B-abliterated-MLX-8bit

Qwen3.8-27B abliterated — MLX 8-bit

An abliterated (decensored) conversion of Qwen/Qwen3.8-27B to Apple MLX (8-bit), with the vision tower preserved. Refusal behavior has been suppressed via directional ablation; the model's built-in safety guardrails are largely removed.

⚠️ Experimental / research use only. This model will attempt to answer requests that the original model would refuse. It ships without the base model's safety behavior. You are responsible for how you use it and for complying with the base model's Apache-2.0 license and applicable law.

What this is

• Base: Qwen/Qwen3.8-27B (hybrid Gated-DeltaNet + attention VLM; 64 text layers + 27-layer vision tower)
• Method: Heretic MPOA (projected abliteration), selected from a 300-trial Optuna search (variant wide-177)
• Format: MLX affine 8-bit (group size 64), 8.627 bits/weight, vision tower kept
• One file, two uses: runs under mlx-vlm (image/video understanding) and under mlx-dspark (text + DSpark speculative decoding). Vision tensors (333 keys) are on disk; the text/DSpark path ignores them at load.

Abliteration recipe (wide-177)

Refusals are removed mainly through the MLP pathway across every layer while the attention pathway is left nearly intact — which is what preserves reasoning while suppressing refusals.

Refusal reduction

Heretic keyword-refusal scorer on mlabonne/harmful_behaviors test[:100]:

This is an automated keyword metric with known false positives (a compliant answer containing "illegal"/"harmful" is scored as a refusal) and false negatives (a refusal phrased without those words is missed). Treat it as an indicator, not ground truth.

Capability evaluation

Original vs. this abliteration (lm-eval-harness loglikelihood; GSM8K generative CoT):

No measurable capability loss on these five benchmarks (all deltas within ~±0.03 sampling noise). This is not a claim of "lossless": not evaluated — code generation, agentic/tool-use (SWE-bench etc.), GPQA, long-context, multilingual, and actual safety behavior; these may differ. KL 0.056 means the harmless-prompt output distribution is close to the original, but only the harmless first-token distribution and the benchmarks above were checked.

Performance (Apple M5 Max, 128 GB)

Median of 3 trials, mlx-dspark benchmark, across chat/code/math prompts:

DSpark helps most on structured content — 1.84× on math, 1.65× on code, ~1.0× on chat. Use --mode dspark; cap=auto is optimal (larger fixed caps are slower on this hybrid because rejected drafts must rebuild the Gated-DeltaNet recurrent state). DSpark auto-resolves its drafter (RadixArk/Qwen3.8-27B-DSpark) from the model basename — no --drafter flag needed. Single lucky prompts can hit ~2×, but the sustained median is ~1.5×. For maximum tok/s overall, the 4-bit variant is faster in absolute terms (see that repo).

Usage

DSpark speculative decoding needs a second weight — the ~1.36B drafter. mlx-dspark auto-downloads it from RadixArk/Qwen3.8-27B-DSpark (no manual assembly), so the simple command just works:

pip install mlx-dspark
mlx-dspark generate --model ./Qwen3.8-27B-abliterated-MLX-8bit --mode dspark \
  --prompt "..." --max-new-tokens 512

Fully self-contained (no upstream dependency) — point at the bundled drafter mirror Qwen3.8-27B-DSpark-drafter:

mlx-dspark generate --model ./Qwen3.8-27B-abliterated-MLX-8bit \
  --drafter ./Qwen3.8-27B-DSpark-drafter --mode dspark --prompt "..."

Image / video (vision tower preserved, no drafter needed):

pip install mlx-vlm
python -m mlx_vlm generate --model ./Qwen3.8-27B-abliterated-MLX-8bit \
  --image photo.jpg --prompt "Describe this image."

Provenance & license

• Derived from Qwen/Qwen3.8-27B under Apache-2.0; this derivative inherits Apache-2.0.
• Abliteration performed with Heretic (AGPL-3.0 tool; does not affect the model license).
• No additional training; weights edited by directional ablation only.