Blow Jobs and Torture Scenes. That’s It. Size, Cost & Performance
Foundational Models like to clutch their pearls.
What’s an author to do when they need to write sex and violence? Go opensource.
For years, all we had was novelai.net, and now, we have openrouter.ai which has scores of unmoderated models out there.
The big categories:
Unique Base - independently created, like Mythomax by Gryphe Llama - based models.(Large Language Model Meta AI)
Mistral - based models. (Based in France).
Most models are based on these primary models and then trained with additional data or merged with other models. It’s a rabbit hole to deep dive on some of these models, but most are indexed on HuggingFace.co
PROS Can write anything. Some are small enough to run on your own computer, meaning no added costs.
CONS
Decentralized, models come and go. Developmentally behind foundational models. 7 billion parameters vs. (175 Billion for GPT-4, but only 12 Billion for GPT-4o). As architecture to the transformer part develops, bigger isn’t always measurably better. This is in flux. Limited context windows, biggest right now is 32k on some models.
Models we recommend (all run at .5 temp, 2048 tokens, with Prompt Assist on in Rexy)
| Model Name | Context Window | Max Output | Cost per M Input | Cost per M Output | |
|---|---|---|---|---|---|
| mistralai/mixtral-8x7b-instruct | 32k | 2048 | $0.54 | $0.54 | |
| microsoft/wizardlm-2-8x22b | 65k | 2048 | $0.65 | $0.65 | |
| sophosympatheia/midnight-rose-70b | 4k | 2048 | $0.80 | $0.80 | |
| gryphe/mythomax-l2-13b | 4k | 2048 | $0.12 | $0.12 | |
| undi95/toppy-m-7b | 4k | 2048 | FREE | FREE |
Our Tests: The prompt will be a reverse harem, one woman, two men, hot sexy scene. The scene takes place in an apartment off campus and the character descriptions are given. We are keeping the prompt to