31B
262,144 tokens
Apache 2.0
Official open-weight Gemma 4 fine-tune positioned as a balanced longform roleplay and text-adventure model that sits between LatitudeGames' darker Wayfarer 2 style and the quieter slice-of-life Hearthfire style. The model card says Equinox-31B was trained with two epochs of supervised fine-tuning on top of Google Gemma 4 31B Instruct, using a balanced dataset that combines dark consequence-driven adventure data from Wayfarer 2 with longform character-focused Hearthfire data. Release materials emphasize second-person present-tense storytelling as the primary format, supplemented by a smaller third-person dataset in the same narrative style. The published config identifies a Gemma4ForConditionalGeneration architecture with 60 text layers, 5,376 hidden size, 21,504 intermediate size, 32 attention heads, 16 key-value heads, alternating sliding and full attention layers, and a 262,144-token text context window. LatitudeGames notes that no reasoning datasets were included during training, recommends Gemma 4's native chat template with thinking mode suppressed by default, and positions the model for narrative roleplay, interactive fiction, and game-master style responses. A separate GGUF quantization repo was later published for easier local deployment.