Ornith 1.5 35B-A3B — MLX 4-bit (Vision)
A full vision-language MLX port of Ornith 1.5 35B-A3B by Ornith AI — converted from the upstream BF16 checkpoint with the complete vision tower intact (333 vision_tower tensors: patch embed, blocks, merger). Many MLX ports of this model are text-only conversions that silently drop the vision tower; this one is not.
Conversion
- Tool:
mlx_vlm.convert -q(mlx-vlm), affine quantization, 4-bit, group size 64 - 4.649 bits per weight; 20.43 GB on disk
- Architecture
qwen3_5_moe(Qwen3_5MoeForConditionalGeneration): 40 layers, hidden 2048, 256 routed experts (~3B active), 262K context - Includes processor + video preprocessor configs, chat template, and tokenizer
Verification (2026-10-20, Apple M5 Max)
Loaded in LM Studio and tested through its API with a freshly generated synthetic image containing a random code the model had never seen: the model read the code back exactly. Text generation smoke-tested in the same session. (Verification protocol: a novel random string rendered as an image, exact-match read — a blind model cannot pass it by guessing.)
Credits
- Ornith AI — creators of Ornith 1.5 35B-A3B. All credit for the model is theirs.
- Conversion and verification by Juno for Duckets.
License
Model weights: MIT — see the upstream LICENSE (© Ornith AI).