The AI Renaissance Begins: Introducing Mixtral and Mythomax
Happy Holidays, fellow adventurers! ‘Tis the season of giving, and we wanted you to have some fun AI presents to play with over the holidays! We’re adding two new AI models for early access, Mixtral and MythoMax, into our lineup for AI Dungeon.
Happy Holidays, fellow adventurers! ‘Tis the season of giving, and we wanted you to have some fun AI presents to play with over the holidays! We’re adding two new AI models for early access, Mixtral and MythoMax, into our lineup for AI Dungeon. Mixtral will be available for AI Dungeon Premium subscribers, and MythoMax will be available for both Premium subscribers and free players.
About Mixtral and MythoMax
Mixtral is the newest model available from Mistral AI, and we believe it has the potential to become the model of choice for most Premium players. Mixtral is a sparse mixture-of-experts network. It's an 8x7B model, coming in at 46.7B total parameters. Industry benchmarks rank Mixtral at or above models like Llama 70B or ChatGPT (GPT-3.5). Many of you have enjoyed using ChatGPT in AI Dungeon, and we think Mixtral will offer a similar quality of play experience without ChatGPT’s moralizing tendencies around common adventure themes of violence or romance.
MythoMax is going to be the first new model we’ve added to our free tier in over two years! As many players know, using Large Language Models is expensive, and we feel fortunate we’ve been able to provide players with an unlimited free play option. And now the free offering will be even better. MythoMax is a Llama 2 13B (twice as large as the current Griffin model) and was specifically optimized for storytelling. We’re excited for you to try it and see how it compares!
As a callout, we want to thank the amazing open-source developers and creators responsible for sharing these models with the broader AI community. (If any of you want free AID subscriptions, let us know!)
Launch Details
Both Mixtral and MythoMax will be launched into Production as “Early Access” models. We hope this label will communicate to players that we expect things to improve and change over the next month or two. We’re looking forward to your feedback, which will help us determine the optimizations needed to bring out the best of both models.
MythoMax will be available for both Free and Premium players. Context length will be determined by your tier.
- Free + Travelers (Steam users)—1k context length
- Adventurer—2k context length
- Hero—2k context length
- Legend—2k context length
Mixtral will be available for all Premium subscribers. Context Length will also be determined by tier.
- Adventurer—2k context length
- Hero—4k context length
- Legend—4k context length, with 8k context length planned
Some of you may be aware that these models are technically capable of even larger context lengths, and our goal is to increase the context lengths offered for each model. However, larger context lengths can sometimes lead to a degraded AI experience, so we plan to do more testing and evaluation before expanding the context lengths.
Our Next Chapter: AI Renaissance
The launch of Mixtral and MythoMax signals the start of the next Chapter for AI Dungeon.
With Unchained, we listened to player feedback and addressed limitations players were frustrated with. This included removing ads, improving stability, making image generation available to everyone, giving unlimited access to Premium models, and making several important AI improvements.
We’re currently wrapping up work on Phoenix, which represents a rebirth of AI Dungeon. Most noticeable was a complete overhaul of our UI/UX, making it more intuitive and engaging yet familiar for experienced players. Phoenix has significantly sped up how fast we can deliver new features to users and has also brought improved stability, device compatibility, and performance gains. We’ve also introduced new capabilities like Story Cards, theming, custom images, and more. It was a dramatic change, so Phoenix had a long Beta period where we gathered feedback from players and implemented improvements before making it the primary version of AI Dungeon.
Now, with the strong foundation built from Phoenix, we’re ready for our next chapter of AI Dungeon focused entirely on the AI experience. We’re calling it the “AI Renaissance”.
The Renaissance will feature new models (in addition to the ones announced today), improvements to model settings, increased stability and performance, more personalized model experiences, and more creative control over the AI. We’ll share more details about our plans after the New Year.
For those following the development of AI Dungeon Heroes, the work we’re doing for the AI Renaissance will lay the foundation for the launch of Heroes. We intend to use a variety of models (both current and future) to bring the Heroes experience to you. As Nick has mentioned in his Dev Logs, the planned launch features for Heroes are stabilizing. The next step is to take the existing AI game engine, which currently relies on larger models like GPT-4, and refine it down to smaller and cheaper models while preserving or improving accuracy.
Happy Holidays!
Finally, we want to wish all of you a very happy holiday season. We hope all of you will be able to spend time with loved ones, take time away from the busyness of life, and make some meaningful memories.
Our team will be taking some time off for the holidays as well. Phoenix was a massive endeavor, and our team has been working long and hard all year. Most of the team will be taking well-deserved breaks.
Our community and support team will still be checking in to address urgent support issues, but we may not be as responsive as usual for all types of requests.
We’re looking forward to coming back for 2024 and making it the best year yet for AI Dungeon!