Models

Since our launch in June 2021, we have released a multitude of groundbreaking storytelling AI models and more! Each of our models bears a mythological namesake and handcrafted artwork to accompany them, this page is dedicated to showcasing their stories.

Current

Erato

/ɛrˈato/

Released on September 23rd, 2024, Erato is based on the Llama 3 70B Base model, with continued training on the most high-quality and updated parts of our Nerdstash pretraining dataset for hundreds of billions of tokens, spending more compute than what went into pretraining Kayra from scratch. We finetuned her with our updated storytelling dataset, tailoring her specifically to the task at hand: telling stories. Early on, we experimented with replacing the tokenizer with our own Nerdstash V2 tokenizer, but in the end we decided to keep using the Llama 3 tokenizer, because it offers a higher compression ratio, allowing you to fit more of your story into the available context.

Kayra

/kaɪrə/

Our in-house, state of the art model, Kayra is the answer to any capability question you might have for NovelAI. With 13 billion parameters of brainpower and trained on roughly 1.6 trillion tokens of data, he can do it all; high-quality storytelling, text adventure, instruct, prose augmentation, and a myriad of corner cases like item lists and video transcripts. In addition, Kayra is equipped with an 8,192-token context window—placing his powers of recall on par with Clio and 4x higher than all our prior models. Kayra was released on July 28th, 2023, with his V1.1 update on August 15th, 2023. Kayra's title is taken from the Turkic creator god of the same name, with additional design motifs for his portrait that pay homage to our special cluster of H100 GPU's, known as 'Shoggy'.

Clio

/ˈklaɪoʊ/

The first of our 'from-scratch' models trained on the Shoggy GPU cluster, Clio represents an incredible amount of knowledge and power condensed into a relatively small model. At 3 Billion parameters and approximately 1.6 trillion tokens for training, what Clio lacks in size she makes up for in speed. Just like Kayra, she also has access to text adventure, instruct, and prose augmentation features via Special Modules, in addition to the same 8,192-token context window. Clio was released on May 23rd, 2023. Clio is named after the Muse of history in Greek mythology, and shares Shoggy-inspired design cues with Kayra since she was also trained on our H100 cluster.

Legacy

Krake

/kreɪk/

Previously one of the largest models to ever have the NovelAI fine-tune applied to it, Krake utilizes EleutherAI's Neo-X 20b model to create a storytelling experience with unique flair. Krake is an exceptional jack-of-all-trades, but has some notable limitations such as slower generation speed, a 2048 token context window, and incompatibility with Modules v2. We no longer recommend Krake for any use case other than personal preference. Krake was released on March 11th, 2022, with his V2 update on April 29th, 2022. Krake's name and design motifs are inspired by the Kraken, a sea monster of Scandinavian mythology.

Euterpe

/juːˈtɜːrpiː/

Based on the Fairseq family's 13b transformer model, Euterpe represents a major leap forward in the history of NovelAI. Though now outdated, Euterpe's compatibility with Modules V2, impressive generation speed, 2048-token context window, and strong adherence to the NovelAI fine-tune made her a marvel for her time in the spotlight. She is one of two models to support Custom Modules, alongside her older brother, Sigurd. Euterpe's experimental release was on January 9th, 2022, followed by her v1 release on January 11th, and her v2 update on February 6th. Euterpe's name and design come from the Muse of music in Greek mythology.

Sigurd

/si:'gʊrd/

Born from the use of the NAI fine-tune on EleutherAI's GPT-J model, Sigurd was NovelAI's second-ever model. Even still, he boasts 6 Billion parameters, 300 million tokens of training time, a 2048-token context window, support for both Modules V2, Custom Modules, and of course lightning generation speeds thanks to his smaller model size. Sigurd's experimental release was on June 16th, 2021, with his updated release on June 17th, V3 release on June 28th, and V4 release on November 11th. Sigurd's name and design are based on a hero from Germanic legends, famous for slaying the dragon Fáfnir. Sigurd's portrait was done by Klein, and selected in a design contest held shortly after the model's release.

Genji

/gən'dʒiː/

A derivative fine-tune of Euterpe, Genji was an experimental model released in an early attempt at making a NovelAI model which was proficient in Japanese text. Clio and Kayra now surpass Genji's capabilities, and the usage of Genji for Japanese generation is no longer recommended. Genji was released on November 10th, 2022, with his V2 update on February 6th. Genji's name and design are loosely inspired by the Heian-era Japanese novel, The Tale of Genji.

Snek

/snɛk/

Another derivative fine-tune of Euterpe, Snek was an experimental model released in efforts to make a model which was dedicated to writing Python code. These capabilities have since been replaced by other models in the field, and the usage of Snek for producing Python code is no longer recommended. Snek was released on November 10th, 2022. Snek's name and design are based on the logo of the Python programming language.

Retired

Calliope

/kəˈlaɪ.əpi/

NovelAI's very first model, Calliope is where it all began. Though quick on her feet and home to 2.7 billion parameters, she lacked most of the quality-of-life tweaks and additional features we've come to enjoy today, and has since been retired from usage. Calliope was released alongside NovelAI on July 16th, 2021. Calliope's name and design are based on the Muse of eloquence and epic poetry from Greek mythology.

Though no longer available in the model selection list, we have released the weights of her model under the GPL-2.0 license in a blogpost here.