Welcome NovelAI’s Newest Writing Model: Xialong

Our newest and most capable writing model yet is finally available for our Opus subscribers. Welcome Xialong!

Xialong is a finetune of ZhipuAI’s GLM-4.6 model using our refined storytelling dataset to bring a much enhanced level of creativity. We kept up the high levels of knowledge and coherence of the underlying base model while bringing the capability to better adapt to your writing style.
Compared to the original and other models trained with RL, Xialong doesn’t have a collapsed token probability space. This means that retries can get different outcomes and also extends to name diversity across all stories.

This finetune was our most ambitious yet, with the model being the largest we’ve worked with at 355 billion parameters in addition to being the first one we’ve trained based on a Mixture of Expert architecture. We are very happy with the end result and hope you will enjoy writing alongside Xialong.

We decided to honor the base model’s Chinese origin by making the persona of the model an ancient Chinese dragon named Xialong (夏龍). The character art was done by Aini, the artist who rendered most of our other models.

Like all other NovelAI trained models, Xialong supports the special formatting you all are used to, such as the ATTG line in front of the story or the memory box, lore entries with a leading “ — — ” line, en spaces for quotes, and many more. We do encourage the use of ATTG at the top of the context, as it helps focus the model

A common and irritating issue was repetitions and loops, where models would repeat text from earlier in the story or even the content of lorebook entries or memory. Xialong was finetuned not only on our secret sauce, but also a flavor of reinforcement learning to teach it not to repeat itself or go into loops. This makes it much more pleasant to use to the point where we no longer need repetition penalty in our presets

One notable difference from past models is that text adventures and text adventure mode now don’t send “>” at the beginning of user input anymore. Instead user inputs are sent as instructions. If you are running text adventures outside of the dedicated text adventure mode, please use the “{“ instruct syntax instead of “>”.

This finetuned model with our anti-repetition training is so strong that it works well with a single default preset out of the box with a lower temperature (0.85). We expect the community to come out with their own presets, but new users should no longer be required to use them to have a good experience. Additionally, we do not recommend using the prefill for this model and have hidden away the input box for it by default. You still have the option to set a system prompt, but you should see if the default system prompt works well for you first.

Xialong is available as an exclusive release for our Opus subscribers right now. There are no immediate plans for a release to lower tiers at this time.

Go forth and explore new horizons with Xialong now!