What is Equinox 31B?
Equinox 31B is a text-adventure and interactive-fiction model from Latitude, the company behind AI Dungeon. It was posted to Hugging Face on 18 May 2026 with GGUF quants, and runs in AI Dungeon on paid plans. Latitude built it with the finetuner Gryphe Padar.
It is Google's Gemma 4 31B Instruct with two epochs of supervised fine-tuning on a blend of two of Latitude's earlier datasets: Wayfarer 2's dark, consequence-driven adventures and Hearthfire's slow, character-led slice-of-life writing. The data is mostly second-person present tense ("You open the door"). No reasoning data was used and Gemma's thinking mode is switched off in the template.
The weights are Apache 2.0. Latitude recommends temperature 0.8, repetition penalty 1.05 and min_p 0.025.
Is Equinox 31B worth it?
Equinox 31B is the strongest choice here for anyone who needs to know where a role-play model came from. Latitude has run AI Dungeon since 2019, the base is Google's Gemma 4 31B, the training recipe is described on the card, and the weights are Apache 2.0, so a studio can ship it in a commercial game without a custom licence.
The design is a compromise on purpose. Latitude's earlier models split the job: Wayfarer 2 for grim adventures where the player can die, Hearthfire for quiet scenes where nothing is forced to happen. Equinox trains on both, so one model can run a dungeon and a candlelit conversation. The card says plainly that if you want the extremes, the specialist models still do them better.
The catch: there is no published benchmark, no human evaluation and no comparison with other models. Adoption outside AI Dungeon is modest, with 68 likes on the main repository and 2,313 GGUF downloads in the last month. Inside AI Dungeon, context is rationed by plan: Arcanum's unofficial guide lists 4K to 6K tokens on the $14.99 Journey plan and 32K to 36K on the $99.99 Ultimate plan, against the 256K the Gemma 4 base supports. Latitude does not publish those limits itself. Running it yourself needs a GPU that can hold a 31B model.
Our take: for a game studio, Equinox is the one role-play finetune a lawyer will sign off on; for a player, AI Dungeon's mid-tier plans are the easier way in, and the context you get matters more than the model you pick.
Where it earns its place
A five-person studio making a narrative game
Ship a game-master model inside the product without licence risk
Self-host the Apache 2.0 weights behind the game's server and prompt in second person present tense, as the card shows; the only cost is the GPU.
An AI Dungeon player who writes long campaigns
Keep a months-long adventure coherent
Equinox on the $49.99 Mythic plan gets 16K to 20K of context according to Arcanum's guide, enough for several sessions of history before the story memory has to do the work.
A researcher studying narrative models
Compare a domain finetune against its base
Run Equinox and Gemma 4 31B Instruct on the same prompts; the card's recipe (two epochs of SFT, no reasoning data) makes the difference attributable.
Before you commit
- Setup
- Try it in AI Dungeon on a paid plan, or download the GGUF from LatitudeGames/Equinox-31B-GGUF and use Gemma 4's chat template with the card's sampler settings.
- Integrations
- Any Gemma 4-compatible runtime (llama.cpp, vLLM, Transformers); AI Dungeon and its multiplayer stories.
- Data and privacy
- Self-hosted, nothing leaves your servers. In AI Dungeon, your stories are processed under Latitude's terms.
- Cost at scale
- Self-hosted: one GPU large enough for a 31B model per serving instance. AI Dungeon: $14.99 to $99.99 per player per month.
- Limits
- No benchmarks; context capped by AI Dungeon plan; no thinking mode; English, mostly second-person writing.
What’s good
- Apache 2.0 weights from an established game company
- Training recipe and data mix described on the card
- One model covers dark adventure and slow character scenes
- Runs in AI Dungeon with no setup
What’s not
- No benchmark or independent evaluation
- Needs a GPU that can hold a 31B model to self-host
- AI Dungeon caps context at 4K to 36K depending on plan
- Not on the free AI Dungeon tier
Best for
Game studios and interactive-fiction developers who need a role-play model with a clean commercial licence and a known origin.
Equinox 31B alternatives
Equinox 31B: common questions
- Is Equinox 31B worth it?
- Equinox is the role-play model with the clearest parentage: a real game company, a real product and an Apache licence, held back only by a lack of independent testing. We rate it 3.9 out of 5.
- How much does Equinox 31B cost?
- Free weights under Apache 2.0. In AI Dungeon it needs a paid plan, from $14.99/month (Journey) to $99.99/month (Ultimate).
- What are the downsides of Equinox 31B?
- No benchmark or independent evaluation; needs a GPU that can hold a 31B model to self-host; aI Dungeon caps context at 4K to 36K depending on plan; not on the free AI Dungeon tier.
- Who is Equinox 31B best for?
- Game studios and interactive-fiction developers who need a role-play model with a clean commercial licence and a known origin.
- How much does Equinox 31B cost for a team or in production?
- Self-hosted: one GPU large enough for a 31B model per serving instance. AI Dungeon: $14.99 to $99.99 per player per month.
- Is Equinox 31B safe for company data?
- Self-hosted, nothing leaves your servers. In AI Dungeon, your stories are processed under Latitude's terms.
- What are the best Equinox 31B alternatives?
- The closest alternatives we have reviewed are Cydonia 24B v4.3, Gemma 4 31B StyleTune, Aion 3.5 and Gemma 4.
Answers checked Nov 7, 2026.