AI Dungeon does not run on one artificial intelligence. It is a text adventure platform from Latitude that began on OpenAI's GPT-2 and GPT-3 and now offers a changing menu of language models, many of them Latitude fine-tunes of open-weight base models from Mistral, Meta, and Google, plus models from DeepSeek and the GLM line published by zai-org. The roster below reflects the game's own help page as checked on October 27, 2026, and it will not stay fixed.
Key takeaways
- The game started as a hackathon project and used OpenAI's GPT-2, then GPT-3, as its text generator.
- Its current help page lists models built on Mistral NeMo and Mistral Small, Llama 3.3, and Gemma 4, along with DeepSeek and GLM models.
- Latitude released Wayfarer-12B as open weights under the Apache 2.0 license, so that model can be run locally.
- The page does not say where the models are hosted, and some "Dynamic" options rotate among undisclosed models.
- Deprecated and removed models appear on the same page, so any list of what the game "uses" has a date attached.
From a hackathon to GPT-3
News coverage reported on February 4, 2021, when Latitude announced USD 3.3 million in seed funding led by NFX, that AI Dungeon began as a hackathon project by Nick Walton and was helped along by improvements in OpenAI's language generator. Walton told the outlet that early output held together at the sentence level but not across a paragraph, that GPT-2 made a clear difference, and that GPT-3 improved it again. GPT-3, described in a paper submitted to arXiv on May 28, 2020, had 175 billion parameters. A game that used it as a backend also depended on the provider's terms. In an April 30, 2021 report, The Register wrote that Latitude added a content filter after OpenAI asked for changes.
What the help page lists in 2026
Latitude's Wayfarer-12B model card shows the pattern the company uses: take a base model and fine-tune it for adventure play. The in-game help page "AI Models and their Differences" applies the same pattern across its list. A sample of the entries, as listed on October 27, 2026:
| Model | Built on | Where it appears on the page |
|---|---|---|
| Muse (12B) | Fine-tune of Mistral-Nemo-Base-2407 | Free-tier list |
| Harbinger (24B) | Fine-tune of Mistral-Small-3.1-24B-Instruct-2503 | Free-tier list |
| Shimmer (26B, 4B active) | Fine-tune of Gemma 4 26B | Free-tier list |
| Nova (70B) | In-house fine-tune of Llama-3.3-70B-Instruct | Paid-tier list |
| Wayfarer Large (70B) | Fine-tune of Llama-3.3-70B-Instruct | Paid-tier list |
| DeepSeek V3.1 and V4 variants | DeepSeek AI models | Free and paid lists |
| GLM 5.x variants | Models from zai-org | Paid-tier list |
| Dynamic Small, Dynamic Large | Rotate among undisclosed models | Free and paid lists |
The page gives context lengths by subscription tier, from 4K on the free tier for Muse to 32K at higher tiers, and knowledge cutoffs for some models only: October 2023 for Harbinger, December 2023 for Nova, April 2024 for Muse, and January 2025 for the Gemma-based Equinox. In plain terms, a story set in 2025 can run into a model that has never heard of it.
Wayfarer: the model anyone can download
Wayfarer-12B is the one model in the lineup whose recipe is public. Its card names Mistral-Nemo-Base-2407 as the base, gives the license as Apache 2.0, and describes a two-stage training run: 180,000 chat-formatted instruct examples first, then an even mix of synthetic adventures and role-play. The adventures came from simulated playthroughs of published AI Dungeon scenarios, with two language models playing narrator and player, each run lasting up to 8,000 tokens or until the main character died. That design is deliberate. Latitude says it wanted players to face real danger and failure, which many aligned chat models avoid. A larger Wayfarer model was trained on Llama 3.3 70B Instruct and is released under Meta's Llama 3.3 license.
Why the answer keeps changing
The same help page carries sections for deprecated models, such as Hermes 3 405B, and removed ones, such as Mistral Large 2 and a GLM 4.6 variant. Models leave when a better base becomes available, when costs shift, or when a provider changes terms. Anyone writing about AI Dungeon, or deciding which setting to pick for a story, is better served by the current page than by a forum post from two years ago. Three claims could not be settled from primary sources and are left out: the exact dates of the 2020 and 2021 model changes, any period on a third-party commercial model other than OpenAI's, and whether Latitude hosts the models itself.
How to check what a game or app runs on
- Find the vendor's own model page and note its date, since rosters change.
- Look for a base model name and a license, which separate a fine-tune from a wholly new model.
- Check whether the app says "Dynamic," "auto," or "smart," which usually means the model is chosen for the user.
- Read the context length, because it limits how much of a long story the model can see at once.
- Compare knowledge cutoffs if the story depends on recent events.
FAQ
Does AI Dungeon use ChatGPT?
The current help page lists no OpenAI models. The game used OpenAI's GPT-2 and GPT-3 earlier, and its present options are built on models from Mistral, Meta, Google, DeepSeek, and the GLM line from zai-org.
Is the AI in AI Dungeon open source?
Partly. Wayfarer-12B is released with open weights under Apache 2.0. Several other options are fine-tunes of open-weight bases but are not individually published on the model card, and some depend on third-party models.
Why do some models forget details sooner than others?
Each model has a context length, the amount of text it can read at once, and the help page lists it by subscription tier. Story Cards and memory settings exist to carry key facts when a story outgrows that window.
Which model is the best one for a story?
No source reviewed gives a reliable ranking. Latitude describes each model's style, and the differences reflect the base model, the fine-tuning data, and the context length rather than a single quality score.
Related reading
- Who builds AI systems and what the job data show
- How much mathematics AI models and their users need
- Which AI tools Canadian businesses report using
- How people use AI chatbots for news
More in AI Tools and Industry.
Last verified: October 2026
Sources
- AI Models and their Differences, AI Dungeon Help Center, Latitude, checked October 27, 2026
- Wayfarer-12B model card and Wayfarer-Large-70B-Llama-3.3 model card, Latitude, Hugging Face
- Language Models are Few-Shot Learners, Brown et al., arXiv:2005.14165, May 28, 2020
- AI Dungeon-maker Latitude raises USD 3.3M to build games with infinite story possibilities, February 4, 2021
- AI text adventure generator report on content filter vulnerabilities, The Register, April 30, 2021