sophosympatheia/Glistening-Gem-31B-v2.0
Glistening-Gem-31B-v2.0 by sophosympatheia is a 31 billion parameter merged language model based on the Gemma-4 architecture. This model is a creative merge of four distinct models, designed to be more stable and arguably more creative than its predecessor. It is intended for creative tasks, though a known issue with its lm_head weights affects raw transformers and newly built quants, making older GGUF quants potentially more stable.
Loading preview...
Overview
Glistening-Gem-31B-v2.0 is a 31 billion parameter language model developed by sophosympatheia. It is a creative merge of four different Gemma-4 based models: TheDrummer/Artemis-31B-v1, zerofata/G4-MeroMero-v2-31B, densenet/Gemma-4-31B-StyleTune-heretic-ara, and llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic. This version (v2.0) aims to improve stability and creativity compared to its v1.0 counterpart.
Known Issues
Important Note: This release has a known problem with its lm_head weights. The fine-tuned lm_head from densenet/Gemma-4-31B-StyleTune-heretic-ara does not merge well, leading to noticeably worse and often incoherent output when using the raw transformers weights or newly created quants. The developer recommends skipping this model in favor of sophosympatheia/Glistening-Gem-31B-v2.1.
Interestingly, GGUF quants made before a specific patch (commit 12ceea82bb63f34c82ecab0626c51173e3991d6e) are often more stable because they inadvertently used a tied token embedding matrix as the output layer, bypassing the problematic lm_head. Users with existing, well-performing GGUF quants are advised to continue using them.
Sampler and Prompting Tips
The model includes recommended sampler settings (e.g., Temp 0.8, Min-P 0.1, Adaptive-P Target 0.6) and a master import JSON file (Glistening-Gem-v2-ST-Import.json) for tools like SillyTavern, providing a starting point for creative tasks. The merge was performed using the DELLA method with densenet/Gemma-4-31B-StyleTune-heretic-ara as the base.