NovelAI - Models

Since the launch of NovelAI in June 2021, the creators have introduced many innovative storytelling AI models and more. Each name of the model is inspired by mythology and handcrafted artwork to accompany them.

Calliope is the first ever model of NovelAI that is released on 16 July 2021. The model has 2.7 billion parameters but lacks in quality and additional features that we use today. Well, this model is no longer available.

Current Models

Some of the current models in NovelAI include −

1. Erato

Erato is the latest model that is based on the Llama 3 70B Base Model that is released on September 23, 2024. The model is continued to train on the updated parts of the Nerdstash pre-training dataset for tokens. The model is carefully tailored, especially for the task of storytelling. To allow the user to fit more of their story into the context, the Llama 3 tokenizer is used, replacing the Nerdstash V2 tokenizer.

2. Kayra

Kayra is released on 28th July, 2023, which is one of the featured advanced models consisting of 13 billion parameters. The model is popular for its high-quality story telling, text adventures, instructions, and prose augmentation, which help excel in handling a variety of tasks like item lists and video transcripts. The model has an impressive 8,192 token context window, surpassing previous models by four times.

3. Clio

Clio is released on May 23, 2023, which is the first 'from scratch' model that is trained on the Shoggy GPU cluster, with 3 billion parameters and 1.6 trillion tokens for training. Though the model is smaller than other models, it compensates with speed, as it efficiently performs tasks such as text adventure, instruction, and prose augmentation.

4. Legacy Models

Krake is released on March 11, 2022, which is one of the largest models fine-tuned by NovelAI and developed using EleutherAI's Neo-X 20b model for a unique storytelling experience. However, it has slow generation speed, a 2048-token window, and is incompatible with Modules v2. Euterpe is another legacy model that is based on the Fairseq family's 13b transformer model, with impressive generation speed and compatibility with Modules V2.

Sigurd is NovelAI's second model that featutes 6 illion parameters and supports both Custom and Modules V2, also known for its lightning-fast generation. Genji and Snek are experimental models derived from Euterpe, with Genji focusing on text generation in Japanese and Snek on Python coding, but they are no longer recommended for story generation after the introduction of the latest models since they are more capable.