X3I4

I recommend using OpenRouter API. Read my guide below. I am not sponsored by OpenRouter and I doubt they would.

Contact: Discord

#Safety and Privacy Warning

To ensure your safety and privacy with the OpenRouter API, it is important to understand data retention and training policies. Some OpenRouter providers save your prompts for up to 30 days, while others may even train new models on your data. For each model in your request, OpenRouter’s default behavior is to load balance requests across providers, prioritizing price. This is not ideal for the roleplaying community. Fortunately, there is a simple built-in workaround to ensure this never happens.

#Zero Data Retention Policy

OpenRouter has a setting that, when enabled, only allows you to route to providers that have a Zero Data Retention policy (ZDR). Some model may become unavailable after toggling this on, since they don't have any ZDR provider.

#Filtered Providers Because a provider has ZDR doesn't necessarily mean it is unfiltered. Some providers, use external moderation layers in addition to the model’s internal safety training. Even if a model (like Llama 3) is technically capable of answering a controversial prompt, a filtered provider will intercept the response. This result in a "refusal" error or a generic "I cannot fulfill this request" message, even if the model itself didn't have a problem with the prompt.

Examples: Most "Big Tech" providers (Google, OpenAI, Anthropic) and some commercial hosts like Cloudflare.

#Unfiltered Providers

Unfiltered providers pass your prompt directly to the model and return exactly what the model generates, without any "safety" checks. If you use an "Uncensored" model (like Dolphin Mixtral), the provider won't stop it from producing NSFW or controversial output.

#Providers Selection

OpenRouter has a setting that allows you to choose exactly which providers you want to use. You can create a whitelist and a blacklist. Toggle "Always Enforce Allowed" if you want to disable fallback providers.

#Quantization

Quantization reduces model size to fit inside a GPU VRAM while aiming to preserve performance. There are many quantization levels. When a model is too compressed, it becomes dumber. It lose its creativity and is prone to repetitions. Most providers shows the quantization level on the model's page.