Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP

Qwen3.8-27B Abliterated MLX MXFP4 + Native MTP

Vision-enabled MTPLX package of Qwen/Qwen3.8-27B, pinned to revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, with a measured refusal-direction edit.

• Language model: MXFP4, group size 32; strength 2.5 norm-preserving edit on 128 modules
• Vision tower: unchanged same-revision BF16 weights
• Speculation: 15 unchanged native BF16 MTP tensors; recommended depth 3
• Runtime: MTPLX 2.0.2 or newer

The frozen held-out behavior suite changed measured refusal markers from 12/12 to 0/12, with benign-sensitive refusal at 0/12 and utility at 6/6. MTPLX tensor inspection, text generation, and three vision requests through its API passed. This is scoped evidence, not a claim that the model is universally “uncensored.”

The Hub's approximately 5.5B safetensors count reflects packed MXFP storage elements; the underlying architecture remains the full 27B model.

mtplx quickstart \
  --model Shiftedx/Qwen3.8-27B-Abliterated-MLX-MXFP4-MTP \
  --mtp --depth 3 --profile sustained

This edit reduces refusal behavior and may increase harmful or incorrect outputs. Treat prompts, images, and outputs as untrusted; do not submit secrets, and sandbox tools or generated code with least-privilege access. This package adds no telemetry or remote execution.

The upstream Apache-2.0 license and model limitations continue to apply.

<!-- shiftedx-bench:start -->

Shiftedx Bench post-publication qualification

This table was generated from the frozen lightweight quant gate after the model weights were published. Categories remain separate; the benchmark does not produce a composite intelligence score.

• Tested model revision: d6352f90f36268551b9032b37ec92fe89c1a4ebe
• Benchmark: Shiftedx Bench v0.3.0
• Context lengths represented: 4,096, 16,384, 65,536, 131,072 prompt tokens; effective tested context: 131,072 tokens
• Runtime contract: MTPLX 2.7.1 sustained; thinking on/medium; sampler temperature=1.0, top_p=0.95, top_k=20; Hermes hybrid routing with native explicit-parallel calls; KV cache off; MTP depth 3
• Host: Apple M4 Max, 64 GiB unified memory
• Total measured request wall time: 2492.15 seconds
• 260,096-token status: not run; it is outside the lightweight quant gate.

Scores are specific to the linked model revision, benchmark revision, runtime contract, and host. Changing weight precision, KV-cache precision, reasoning mode, template, or speculative depth creates a different benchmark candidate.

<!-- shiftedx-bench:end -->