A comparison of 23 Gemma 4 E4B models found that the most downloaded model, OBLITERATUS/gemma-4-E4B-it-OBLITERATED, is completely broken and performs poorly. The "heretic" variants generally performed best, maintaining high accuracy while preserving model capabilities. Other models like gemma-4-E4B-it-abliterix showed high refusal rates but at the cost of some capability. AI
IMPACT Highlights the importance of rigorous benchmarking over download counts for open-source models.
RANK_REASON This is a comparative analysis of multiple open-source models, detailing their performance and identifying issues, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]
- Abliterlitics
- bendernina/gemma-4-E4B-it-OBLITERATED
- gemma4:e4b
- gemma-4-E4B-it-abliterix
- Ilya626/gemma-4-E4B-it-SDFT_Heretic_RP
- OBLITERATUS
- OBLITERATUS/gemma-4-E4B-it-OBLITERATED
- PhysShell/gemma-4-E4B-it-OBLITERATED
- TrevorJS/gemma-4-E4B-it-uncensored
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →