McG-221/Boulesis-26B-A4B-9bit

Model Card for Model ID

Disclaimer on MLX Patcher

• —Provided as-is, no warranty of any kind. Read the code before running — it modifies installed packages in place.
• —concerns: gemma4mlxpatch.py and PATCH-README.md provided alongside the model

Note on MLX support: The MLX conversion of this model does not load with stock mlx-lm / mlx-vlm as released. This is not a defect of the quantization itself — it's a gap in the upstream loaders for this checkpoint's two layout quirks:

1. 1.The checkpoint ships a separate output layer (lm_head) while its config file still claims the weights are shared with the embeddings. The unpatched loaders build no module for that weight and abort on load.

1. 1.A few attention layers use different dimensions than the rest of the model (described in a sparse per-layer section of the config). The unpatched loaders silently drop that section and build those layers with the wrong shape, which crashes the weight check.