BlueNipples/LM-Head-Experiments_GGUF

So these are just gryphe style LM head swaps, but I found interpolating the new head with the old at about 65-35 ratio preserved most of the base models instruct/reasoning capability unlike a full swap.

I used the original BF16/Safetensors heads from the models below, interpolated them with a tiny python script, and then put the resulting head back onto the GGUF body. The resulting heads are Q8_0.

I found success with:

WorldSim + Melody WorldSim-Opus-3.6-35B-A3B + Melody1437-35B-A3B

HighIQ + Animus Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED + Qwen3.5-9B-Animus-V13.0

Let's call that last one an experimental v2 of LucidVitality.

The whole thing is fairly easily replicable using a tiny python script and Wiself/Voice.