TheDrummer/Fallen-Command-A-111B-v1

111B • Updated Mar 24, 2025 • 28 • 27

Note The darkest, most based model in existence. It has "concise yet deep" output, useful for "establishing foundations" with other models. CONS: Slow as fuck. Use IQ2_XXS for offloading. Fallen Command speaks truth where other models (even uncensored) just repeat bullshit.

coder3101/Skyfall-31B-v4.2-heretic

31B • Updated Jun 2 • 130 • 4

Note Better than any 12B or 24B. Skyfall Heretic completely dominated all other models (GLM/Gemma/etc) in a creative writing competition. It is adept at system prompt following and style emulation. Perhaps it struggles at longer context where GLM/Gemma may outshine it.

Naphula/Ouroboros-24B-v1.4

Text Generation • 24B • Updated about 20 hours ago • 159 • 2

Note Possibly better than Goetia 1.4, more testing is needed to confirm.

Naphula/Goetia-24B-v1.4

Text Generation • 24B • Updated 3 days ago • 1.04k • 8

Note The best 24B, almost as good as Skyfall, but struggles sometimes with long context (due to being a merge). In multiple writing contests, the top 3 were usually 1) Skyfall 2) Goetia 3) GLM

concedo/Huihui-GLM-4.5-Air-abliterated-lossytensors

Text Generation • 110B • Updated Oct 13, 2025 • 24 • 5

Note My #1 local model for dataset augmenting and world knowledge. Excellent at system prompts, right behind Goetia in terms of style adherence. Usually smarter than Gemma 3. The model is 99% uncensored, but sometimes the rare refusal, disclaimer, or ablation artifact leaks through. If 64GB RAM or less use DarkArtsForge/huihui-ai_Huihui-GLM-4.5-Air-abliterated-Q3_K_M-GGUF

mlabonne/gemma-3-27b-it-abliterated

Image-Text-to-Text • 27B • Updated Mar 21, 2025 • 2.68k • • 337

Note Great at system prompts. Seems to have less knowledge than GLM Air but is fully 100% uncensored. When combining them via double pass outputs, GLM->G3 works better than G3->GLM. G3 is better than G4 at pretty much everything IMO, and I have no examples where I can recommend any Gemma 4 over this. It's generally a good fallback model, stable up to 28K.