Favorite Models
The main models kept on my SSD for creative writing, knowledge and RP.
111B • Updated • 166 • 29Note Deepest knowledge despite shorter outputs. Used less often than Cydonia or GLM Air due to its slow speed. However it is still my #1 favorite model of all time and can "set the tone" for other models to continue from. - For my system I use IQ2_XXS.gguf --flashattention --contextsize 8192 --gpulayers 12
coder3101/Skyfall-31B-v4.2-heretic
31B • Updated • 155 • 8Note Excellent writer, one of the best, although it's hit and miss with intelligence. Sometimes it's smarter than Cydonia and GLM Air, other times it's dumber. I recommend using all 3 models and comparing their outputs. Absolutely destroys Gemma 4. Q4_0.gguf --flashattention --contextsize 16384 --gpulayers 19
BeaverAI/Cydonia-24B-v4x-GGUF
24B • Updated • 254 • 5Note Works well with SRP, favorite 24B. Few refusals/disclaimers. Sometimes smarter than GLM Air - Stable up to 128K context, but I recommend using Q8_0 if going past 64K as the speed gains of lower quants become negligible while the intelligence loss sharpens at higher context. Even at 128K it has excellent context recall. Cydonia 4.3 Heretic is too compliant, while original is too censored. v4x doesn't refuse outright yet it has enough pushback to increase the variety and uniqueness of its output.
Naphula/GLM-4.5-Air-Abliterated-FQ3_K_XL-GGUF
Text Generation • 107B • Updated • 210 • 5Note The best uncensored knowledge/speed ratio, works well with SRP and in conjunction with Cydonia. - Stable up to 32K context before hitting pagefile on my system, 16K for fully offloaded MoE
mlabonne/gemma-3-27b-it-abliterated
Image-Text-to-Text • 27B • Updated • 2.94k • 350Note Better than Gemma 4 for uncensored tasks and translations, works well with SRP. Not quite as smart as Mistral or GLM and is prone to making several contextual errors. - Recommended IQ4_XS for 8GB VRAM max supported context 28K --flashattention --contextsize 28672 --gpulayers 14
EldritchLabs/MN-12B-Mag-Mell-R1-Uncensored-Scale1.2
Text Generation • 12B • Updated • 37 • 14Note Favorite 12B, works well for dataset generation, much better prose than Gemma 4. - IQ4_NL fits entirely on VRAM with 4K context or IQ4_XS with 6K