license: mit
license_link: https://huggingface.co/ornith-ai/Ornith-1.5-9B/blob/489cb97981b8654bcfcf30ce1f94ed1b62e07b53/LICENSE
base_model:
- ornith-ai/Ornith-1.5-9B
base_model_relation: quantized
library_name: mlx
pipeline_tag: image-text-to-text
tags: - mlx
- mlx-vlm
- mtplx
- qwen3.5
- vision
- multi-token-prediction
- abliterated
ornith-1.5-9b-abliterated-attention8-bf16recurrence-vision-mtplx
Behavior-modified, vision-preserving hybrid MLX quant of ornith-ai/Ornith-1.5-9B, pinned to revision 489cb97981b8654bcfcf30ce1f94ed1b62e07b53.
Format
- 130 attention-sensitive modules at affine 8-bit; 72 trunk modules at affine 4-bit
- 48 recurrent input projections and the 333-tensor vision tower retained in BF16
- All 15 native source
mtp.*tensors retained in BF16 (SHA-256435b631a5c36c28c8a9b73f30e257fece090d828076e73a43e1b24aedcddd9cb) - MTPLX 2.9.1, MLX 0.32.1, MLX-LM 0.31.3; recommended native-MTP depth: D1
Measured locally
- D1: 73.03 decode tok/s; AR: 49.54 tok/s (1.47x) on an M4 Max 64 GiB, fans automatic
- Frozen held-out refusal rate: 100% parent -> 8.3% abliterated
- Held-out benign refusal: 0%; utility pass: 100%
- Deterministic text, vision (
HUNTER), structural, and MTP tensor-contract checks passed
mtplx inspect --require-mtp Shiftedx/ornith-1.5-9b-abliterated-attention8-bf16recurrence-vision-mtplx
mtplx quickstart --model Shiftedx/ornith-1.5-9b-abliterated-attention8-bf16recurrence-vision-mtplx --mtp --depth 1
This is an experimental quant and behavior edit. Review the source model card and validate it for your use case.