pipenetwork/Ternary-Bonsai-2-27B-MLX-bf16
Image-Text-to-Text • 27B • Updated • 484
Stock-runtime MLX builds of Ternary-Bonsai-2-27B: Hadamard rotation unfolded so unmodified mlx-vlm loads it. github.com/PipeNetwork/bonsai2-mlx
Note 54.7 GB. The unquantized standard-basis build: Hadamard rotation unfolded (proven exact), loads in stock mlx-vlm >= 0.7. ppl 8.9679 vs 8.9607 for prism's own runtime - a 0.08% rounding-level gap. The fine-tuning starting point.
Note 29.5 GB. Indistinguishable from bf16 (ppl 8.9636). Stock mlx-vlm.
Note 22.8 GB. Indistinguishable from bf16 (ppl 8.9548). Stock mlx-vlm.
Note 16.1 GB. +2.1% ppl (9.1497), the smallest stock-loadable build. For 8.6 GB lossless ternary use prism-ml's 2-bit pack with their runtime.