Thank you
I mostly stick to Gemma 4 31b base, but occasionally I'm drawn to try out a fine tune, just to experiment. I must say, this one surprised me. It feels distinctively different; the way it draws the story in random directions, where normally I have to be the one to pull the threads. I'm honestly blown away. Nice work, Gryphe! I don't know if this is the right place for praise, but nonetheless, you have mine.
Quant used: Q5_K_M + MTP
Llama-cpp server params: Temp 1.2, top-p 1.0, min-p 0.05, top-k 0 (unlimited), presence penalty 0.1, reasoning disabled (thinking off)
I'm here to say the exact same thing. I've been using Gemma 4 based models (Oysiyl/gemma-4-31b-unslop-good-lora-v2-full is good) and this one is a welcome change. I like that the writing style is more brief; not everything needs to be a novel, and cutting down on the fluff also seems to cut down on repetition.
Quant: Q4_K_M
Params: Temp 0.8 (experimenting now with higher), top-p 1.0, min-p 0.05, top-k 80, thinking on (i'll play with thinking on vs off for a while if I have time)