Modelpublic

Shiftedx/ornith-1.5-9b-affine4-bf16recurrence-vision-mlx

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes101downloads

Model Card

ornith-1.5-9b-affine4-bf16recurrence-vision-mlx

Affine 4-bit MLX quant with recurrent-state inputs retained in BF16 and the parent model's BF16 vision tower. Built from Ornith-1.5-9B revision `98db59b` for Apple Silicon.

Format

—Size: 6.54 GB
—Language trunk: 202 affine-4 modules and 48 recurrent-input modules in BF16
—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: `d4efb82`.

Usage

bash

python -m mlx_vlm.generate \
  --model Shiftedx/ornith-1.5-9b-affine4-bf16recurrence-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	5/10	69.0 tok/s	5.61 GB
Long context	3/15	60.9 tok/s	5.61 GB
Tools	6/6	63.1 tok/s	6.71 GB
Agentic	2/2	—	—
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.