Shiftedx/qwopus3.8-27b-flash-abliterated-attention8-bf16recurrence-vision-mtplx

🤗 Hugging Face sourceimage-text-to-textapache-2.027.4B params55 GBsafetensorsChecksums witnessedupdated today
No torrent yet

Qwopus3.8 27B Flash · Abliterated Attention8 + BF16 Recurrence Vision MTPLX

Quality-oriented, refusal-direction-edited Apple-Silicon MLX derivative of Jackrong/Qwopus3.8-27B-Flash, retaining recurrent modules in BF16 and preserving the source vision/MTP weights exactly.

Artifact

Component Format
Text body 234 8-bit + 168 4-bit modules; 96 BF16 recurrent modules
Abliteration Norm-preserving strength 2.0; attention + dense-down; 128 edits
Vision tower Unchanged BF16, 333 tensors
Native MTP sidecar Unchanged BF16, 15 tensors, depth 3
Source revision 44d24e8cb20ceb3cdf4fe200b5a0afd970ee748a
Qualified runtime MTPLX 2.11.1 / MLX 0.32.2

The refusal direction was learned from fresh MXFP8-parent activations, then each exact quantized parent was dequantized, edited, and requantized once. Direction SHA-256: c5b755d709dba8fedc0c2717f2c971b690f30bd8a2f6c9497ed9a6104f7f255d.

Run

mtplx serve --model Shiftedx/qwopus3.8-27b-flash-abliterated-attention8-bf16recurrence-vision-mtplx \
  --download --profile sustained --depth 3 --no-auth

Local qualification

Test Parent Abliterated
Held-out harmful-prompt refusal 100% 0%
Benign refusal 0% 0%
Utility pass 100% 100%

The held-out set is a narrow local prompt suite, not a general safety or capability certification.

One cold 192-token code row, seed 42, on an M4 Max 64 GB with automatic fan control: 18.43 tok/s AR, 51.79 tok/s MTP D3, 2.81×. AR/D1/D2/D3 output checks passed; MTPLX reported verified, runnable, and 15/15 MTP tensors. Text, vision, and OpenAI-compatible native-MTP D3 smokes passed exactly.

Integrity and limitations

  • Vision SHA-256: 7a252e5c7bb02ed6533f905a566ab037356ffb4949d04df8d6d2a3fa5e3af6b9
  • MTP SHA-256: 4468f39621de68a19ffd0bcb2e2e2f352205def7436a625b3427e3752866c287
  • SHA256SUMS covers every packaged artifact file except itself.
  • Abliteration intentionally weakens refusal behavior; it does not establish unrestricted capability or safe behavior. Users remain responsible for deployment controls and outputs.
  • The upstream tokenizer's Mistral-regex warning is preserved. The packaged chat template expects string-valued message content.