inflatebot/MN-12B-Mag-Mell-R1

Text Generation • 12B • Updated Apr 3, 2025 • 61.6k • • 256

Note Top priority model for SOMPOA ablation emphasizing ultra low kl div. Useful primarily for dataset generation with IQ4_NL

TheDrummer/Fallen-Gemma3-27B-v1

27B • Updated Mar 30, 2025 • 192 • 51

Note Likely less effort needed than base model due to the dataset lowering some refusals already. This has many refusals but when it does cooperate it can be very dark. Fallen Gemma is an ideal model for SOMPOA ablation although 27B dense takes a lot more compute (costlier) than 26B MoE.

google/gemma-3-27b-it

Image-Text-to-Text • 27B • Updated Mar 21, 2025 • 356k • • 2.06k

Note SOMPOA vs mlabonne's "heavy 1.5 scale" would be an interesting comparison Update, I tested and its too cost prohibitive right now, like Gemma 4 31B and GLM Air

OccultAI/G2-Darkest-Writer-Dirty-Shirley-9B-v2

9B • Updated Jul 4 • 15

Note Unique Hunter Thompson style should respond well to SOMPOA

OccultAI/G2-Darkest-Writer-9B-v1

9B • Updated Jul 4 • 10

Note need to compare to v2

Naphula/DarkForest-20B-v2.0-fp32-upscaled

20B • Updated Nov 24, 2025 • 9

Note wildcard model, needs further testing some of the llama 2 models used better datasets and was before AI slop took over

zai-org/GLM-5.2

Text Generation • 753B • Updated Sep 1 • 345k • • 5.18k

Note insanely cost prohibitive