This is LFM2 8B A1B upcycled to double its capacity, based on MuXodious' version for low refusal rate. The new experts were fine tuned on a variety of datasets to utilise the increased parameter count effectively, but I recommend using this model as a base for fine tuning, though it is fully functional on its own. I also increased the active parameter count to 2B by increasing top k to 8.
