Standard group-32 MXFP8 language quant with the same-parent BF16 vision tower. Built from ornith-ai/Ornith-1.5-9B revision 98db59b for Apple Silicon.

Format

Logical size: 10.17 GB
Dense Qwen3.5 9B language trunk: 250 uniform MXFP8 group-32 modules, including recurrent inputs
Vision: 333 same-parent BF16 tensors
Context metadata: 262,144 tokens
Runtime: autoregressive MLX-LM/MLX-VLM
Native MTP: unavailable; the upstream checkpoint contains no mtp.* tensors

Exact precision and conversion records are included in BUILD_RECIPE.json and conversion_receipt.json. Immutable weight revision: 915e4f4.

Usage

bash

python -m mlx_vlm.generate \
  --model Shiftedx/ornith-1.5-9b-mxfp8-vision-mlx \
  --image image.jpg \
  --prompt "Describe this image." \
  --max-tokens 256

Qualification

ShiftedX Bench v0.3.0 commit 3bbb0bfa01e33503163cb34ef52b4d507e456265, Apple M4 Max / 64 GiB, thinking enabled, medium reasoning, temperature 1, top-p 0.95, top-k 20, KV cache off:

Lane	Result	Mean decode	Peak active memory
Quality	6/10	46.2 tok/s	9.23 GB
Long context	1/15	42.7 tok/s	9.23 GB
Tools	6/6	43.4 tok/s	10.34 GB
Agentic	1/2	—	—
Vision, native MLX-VLM	4/4 strict	—	—

Text lanes used MTPLX 2.7.1 in stock autoregressive mode. Vision used native MLX-VLM because MTPLX AR rejects image content. Structural loading and deterministic text/vision smokes passed. Quantization can change behavior; see the parent model card for intended use and license details.