Shiftedx/qwopus3.8-27b-flash-attention8-bf16recurrence-vision-mtplx
Qwopus3.8 27B Flash · Attention8 + BF16 Recurrence Vision MTPLX
Quality-oriented mixed-precision Apple-Silicon MLX release of Jackrong/Qwopus3.8-27B-Flash, retaining recurrent modules in BF16 while quantizing attention-heavy paths more aggressively.
Artifact
The conversion used the pinned BF16 safetensors source—not a requantization of the GGUF release.
Run
mtplx serve --model Shiftedx/qwopus3.8-27b-flash-attention8-bf16recurrence-vision-mtplx \
--download --profile sustained --depth 3 --no-authThis starts the local OpenAI-compatible endpoint. The packaged tokenizer, processor, vision weights, chat template, and MTP runtime contract are included.
Local qualification
One cold 192-token code row, seed 42, on an M4 Max 64 GB with automatic fan control:
AR/D1/D2/D3 output checks passed. MTPLX inspection reported verified, runnable, and 15/15 MTP tensors. Strict text load returned the expected 42; vision identified the test shape/color exactly; the OpenAI-compatible native-MTP D3 smoke also returned 42. These are narrow local qualification results, not a cross-hardware benchmark.
Integrity and notes
- Vision SHA-256:
7a252e5c7bb02ed6533f905a566ab037356ffb4949d04df8d6d2a3fa5e3af6b9 - MTP SHA-256:
4468f39621de68a19ffd0bcb2e2e2f352205def7436a625b3427e3752866c287 SHA256SUMScovers every packaged artifact file except itself.- The upstream tokenizer's Mistral-regex warning is preserved for exact source lineage. The packaged chat template expects string-valued message content.
- Quantization can change quality. Apply the source model's license and limitations to this derivative.
