Lambent/

Arsenic-Shahrazad-12B-v4.4.2-mtp-gguf

A version update meant to steer back towards v4.4's voice, using top-32 logits to strongly anchor towards the distribution while leaving the remaining ~2% probability mass available for MTP head co-adaptation.

Approximately matches the performance of v4.4.1's head and model.

Fork at https://github.com/lambentLogic/llama.cpp/tree/mtp-llama-mistral to actually utilize the head, upstream MTP keeps improving and I'll try to keep up with merging those in.