AI Models
Each model has its own voice and strengths. Pick whichever suits the story you're telling — you can switch at any time from the model selector in the editor.
Lunaris
8B · Llama 3Sao10K
A lightweight model built on Meta's Llama 3, created by Sao10K as a merge of several roleplay-focused finetunes. The base is a couple of generations old now, but Lunaris remains a staple on AI writing platforms for its atmospheric prose and quick, expressive dialogue.
Mistral Nemo
12BMistral AI · NVIDIA
A 12B model from Mistral AI, built in collaboration with NVIDIA and released in 2024. A longtime favorite of the roleplaying community: it punched well above its weight at launch and, though no longer new, still sees heavy use today for its reliable, characterful writing.
Muse
12B · Mistral Nemo finetuneLatitude Games · Gryphe
Developed by Latitude Games — the studio behind AI Dungeon — in collaboration with Gryphe Padar. Muse is a finetune of Mistral Nemo with added emotional depth and stronger character development. It was trained exclusively on second-person, present-tense prose (“you see”, “you feel”), so it performs best when you write in that style.
DeepSeek 3.2
Large MoEDeepSeek
A large, capable model from DeepSeek. It excels at creative writing — reading subtext, inventing fresh and exciting situations, and pushing the plot forward. Many writers prefer it over DeepSeek's newer releases for roleplay and storytelling.
Gemma 4
31BGoogle's open Gemma 4 model, in its 31-billion-parameter dense form. Its standout strength is instruction following — it sticks closely to your directions, making it a strong pick when you want the AI to respect specific guidance and stay on track.
MiMo 2.6 Flash
ReasoningXiaomi
A popular model in the roleplaying community despite being quite new. Great at instruction following and character consistency.
Hemmingway
27BHemmingway
Hemmingway-1 is a 27-billion-parameter model trained specifically for writing prose. It excels at emmotional intelligence and sounds a lot more like a person than other larger models.