Confusing Ablated with Local Models
Ablated AI models run on remote servers, not your laptop. When you request a response, your code sends data to a hosted endpoint. Local models like those in Ollama run entirely on your hardware. You must adjust your SDK configuration to point to the correct base URL.
Network latency affects response time. Your local model speed depends on your GPU. Remote API calls add milliseconds for round trips. This is normal for cloud services.
Local models allow full data privacy. Remote APIs require trusting the provider. Ablated AI does not use prompts for training. Your data stays yours. You still need to manage API keys securely.
Context Window Misconceptions
The context window includes both input and output tokens. You have 100,000 tokens total per request. This sum counts the prompt and the completion together.
- Input tokens count toward the limit.
- Output tokens count toward the limit.
- The max output is 16,000 tokens per request.
If you do not set max_tokens, the model defaults to 2,048 tokens. This prevents runaway costs. You must calculate your available space carefully. Long conversations consume the window quickly.
Assuming Zero Refusals Ever
Uncensored does not mean infinite. The model refuses lawful adult content freely. It does not refuse creative writing. There is one hard limit that always applies.
Sexual content involving minors is refused. This is non-negotiable. Other controversial topics pass through. Security research outputs appear raw. You get the full answer.
Refusals are free of charge. You do not pay for tokens that are rejected. This saves money on bad prompts. The model is tuned for adult use cases.
Token Count Confusion
Pricing is based on real token usage. You pay $0.15 per 1M input tokens. You pay $0.50 per 1M output tokens. Errors cost nothing. Refusals cost nothing.
- Input tokens are charged at the lower rate.
- Output tokens are charged at the higher rate.
- Prepaid credit covers all usage.
Credit never expires. You can top up when needed. No monthly fees apply. You only pay for what you consume. This model suits bursty workloads well.
Streaming vs. Non-Streaming Data
Streaming returns data in chunks. Server-Sent Events (SSE) deliver tokens as they generate. You see output in real time. This improves user experience significantly.
Non-streaming waits for the full response. You get all tokens at once. This is simpler to implement. Streaming requires handling partial data.
The last chunk contains token usage stats. You can track costs accurately. Both modes support the same model features. Choose based on your UI needs.
Function Calling Limitations
Function calling works via tools. You define schemas in your request. The model returns structured JSON. This enables complex workflows.
- Use tool_choice to force function calls.
- Supports JSON mode for strict output.
- Parameters like temperature affect reliability.
Not all features are available. No embeddings exist. No image generation exists. No fine-tuning exists. Focus on text completion tasks. The model handles tools well.
Crypto Payment Nuances
Payments require cryptocurrency only. You can use USDT on TRC20. You can use USDC on Base. Mistakes are fixed via support. No phone number needed for signup. Google or email login works.
Technical Implementation Details
Use the OpenAI compatible SDKs. Point to HTTPS://api.ablated.cc/v1. Set your API key in the header. The model id is "uncensored". It is an open-weight model.
Parameters like temperature and top_p work as expected. You can set a seed for reproducibility. Presence penalty and frequency penalty adjust repetition. These features give you control.
- Rate limit is 300 requests per minute.
- Concurrent requests limit is 8 per key.
- Max request body is 8 MB.
