Update: I manually revised the chat template to change "<think>" -> "<think>\n" to fix inconsistent reasoning initialization. New templated included as jinja and embedded in gguf.
Quant optimized for quality / speed on a Strix Halo 128GiB system. Possibly also beneficial on DGX Spark and similar systems.
Left more headroom on this quant to better utilize long ctx and eventual DFlas once supported.
Refer to the tensor types for a breakdown of quantization method.
See the GLM version for more details on theory and comparisons.