The Art Tool now uses a new model
What exactly does the Nomicrafter Art Tool do?
The Art Tool is a simple tool I created that takes your input, sends it to an LLM along with a system prompt that teaches it how to write for Nomi’s Art Tool, and then returns a JSON object that includes the natural language prompt and a categorized list.
What is changing?
I use an inference provider called Groq (No, not the nazi one) to handle this. Recently, Groq has decided to discontinue the model I had been using, Llama 4 Scout 17B.
I had to change the model to a different one they provide, and I chose GPT OSS 120B.
Groq has a limited stable of models for me to choose from, and there really were few options for the task I wanted to do that weren’t going to hit rate limits after ~2 API calls per minute.
Looking ahead 🧑🚀
Scout really was the better model for this tool, but it’s gone now. GPT OSS 120B is the new model, and sometimes it is quirky. Once in awhile it will write the category name with a colon into the final prompt, and also tends to get a bit wordy. Users online also seem to say that GPT OSS 120B has a lot more guardrails baked in, and will tend towards refusal for anything mature/graphic. That shouldn’t effect Nomicrafter too much, considering my tool is designed to create prompts that adhere to Nomi.AI’s rules.
I’ll be monitoring it in the coming weeks to see if I need to make adjustments to the system prompt, max tokens, or temperature.
If GPT OSS 120B ends up not being suitable for this task, I’ll look into using a different inference provider. In my private life, I use OpenRouter via API. If I can find a free model that isn’t too heavily rate limited, I would consider swapping. OpenRouter has 100x the selection of Groq, including some of the brand-spanking-new chinese models (though for this tool I would look for something efficient, stable, mature, and cost effective).
Tell me how you feel.
I would love to hear your feedback on the output of the new model. If you have the time, get in touch with me on the Contact page and share your input, the prompt it gave you back, and how you feel it could be better.
Thanks for reading!