Equinox-31B Weaves Dungeon Crawls Into Candlelit Conversations

    
        By vramkickedin    
     | 
    
            June 1, 2026 at 10:22 am        
    
     | 
    
        2 min read

Equinox-31B is a new storytelling model from LatitudeGames that blends dark, high-stakes adventures with quiet slice-of-life narratives. It was fine-tuned on Google’s Gemma 4 31B Instruct using two distinct creative datasets, producing a versatile AI for text adventures and longform writing. The model writes in second-person present tense, making interactions feel like a gamemaster guiding you through a living story.

LatitudeGames, the team behind AI Dungeon, developed Equinox-31B to give users more narrative range without needing separate specialist models. They combined their existing Wayfarer 2 data (unforgiving survival stories) with new Hearthfire data (emotional, character-driven scenes) over two epochs of supervised fine-tuning. The result is equally comfortable handling dungeon crawls and candlelit conversations, and can be accessed through an AI Dungeon subscription or run locally via GGUF files.

A balanced storytelling model

Key features

Balanced for dark adventures and slice-of-life.
Fine-tuned on Gemma 4 31B Instruct.
Uses Wayfarer 2 and Hearthfire hybrid datasets.
Second-person present tense narrative style.
Thinking mode suppressed in chat template.
Recommended temperature 0.8, repetition penalty 1.05.
Available via AI Dungeon or local GGUF.
Open to community feedback for future models.

Writers and gamemasters looking for a flexible AI to generate interactive fiction will find Equinox-31B useful. Hobbyists with a powerful local GPU can run the 31B model for private, fast storytelling. AI Dungeon subscribers can skip setup and try it immediately through the existing platform.

Developer notes and limitations

LatitudeGames intentionally made Equinox a middle ground, so it won’t deliver the extreme grimness of Wayfarer or the pure slow-burn atmosphere of Hearthfire. No reasoning datasets were included during training, which keeps the focus squarely on narrative style rather than complex logic. The team plans to open-source similar models and will use community feedback to improve behavior in future releases.