Ornith-1.5-122B, 4-bit MLX affine

ornith-ai/Ornith-1.5-122B (revision 346ec829dd63e107aad8771d74e885e901cec95d, a Qwen3.5-MoE 122B-A10B with a vision tower, MIT) quantized with mlx-vlm 0.7.9 to 4-bit affine weights in groups of 64 (routers and shared-expert gates 8-bit, 4.544 bits a weight, 65 GB). The vision tower, the tokenizer and the chat template are the release's.

Served by TensorFold's native Metal engine on one Apple M5 Ultra by the ornith-15-122b-mlx-m5-ultra recipe, with the incoai/Ornith-1.5-122B-DFlash2 drafter, images, a prompt cache and up to four streams; the recipe's README has the measurements. mlx-vlm loads it as well.

Conversion: python -m mlx_vlm.convert --hf-path <the release> --mlx-path . -q --q-bits 4 --q-group-size 64 --q-mode affine --dtype bfloat16.