FrenchCastle/Solomon-MLX-8bit

Card: runs at BF16 speed (runtime-bound, 17 s prefill + 2 s per branch), 36 GB memory, MLX cache cap in Quickstart, 4-bit build marked in preparation

model card: full open-source card

Solomon ac4f9ccf, quality-q8: 8-bit affine group 64 on the language-model Linear layers (31.6 GB, 25 bound files)

model card and runtime patch

initial commit