Uncensored Artificial Intelligence Production Checklist

Verify model identity and weights

Start by confirming the model serving your requests is the intended uncensored large language model and not a routed third-party service. Send a POST request to GET /v1/models and verify the returned ID is exactly uncensored. This model is an open-weight variant tuned for unrestricted responses, distinct from GPT, Claude, or Llama, and runs on our dedicated servers. Check the weights by injecting a specific instruction in the system prompt that forces a direct answer without hedging. If the response includes standard AI disclaimers or refuses a lawful adult topic, the model identity may be incorrect. Ensure your SDK configuration points to https://api.apiintelligenceartificielle.com/v1 to avoid accidental routing to other endpoints. This verification step prevents integration errors where developers assume they are using our specific uncensored model but are instead hitting a different service.

Test context window boundaries

Validate the 100,000-token context window by sending large inputs and observing retention of early instructions. The limit applies to the sum of prompt and completion tokens. Begin with a prompt of 50,000 tokens and a system instruction at the very start. Request a completion of 16,000 tokens, which is the maximum per request. If you do not set max_tokens, the default is 2,048 tokens. Monitor whether the model remembers the initial instruction after processing the bulk of the context. Context degradation is common in long sequences, so test edge cases near the 100k limit. Record the token usage from the response headers to ensure accurate billing. This test ensures your application can handle lengthy documents or extended conversations without losing coherence or forgetting key constraints.

Validate streaming performance

Test Server-Sent Events (SSE) streaming to ensure low-latency token delivery for real-time applications. Use the stream: true parameter in your request. The final chunk of the stream must contain the complete token usage statistics. Measure the time between token generation and receipt to calculate latency. Streaming reduces perceived wait time but increases complexity in parsing. Ensure your client library handles partial JSON structures correctly if you are not using JSON mode. Compare streaming performance against non-streaming requests to identify any bottlenecks. This validation is critical for chat interfaces where users expect immediate feedback. Confirm that the token counts in the final chunk match the actual tokens sent, ensuring billing accuracy.

Check token usage accuracy

Verify that token counting matches billing to avoid unexpected costs. The API charges $0.15 per 1M input tokens and $0.50 per 1M output tokens. Errors and refusals are free, so ensure your error handling logic does not charge for failed requests. Check the usage object in the response for p_tokens and c_tokens. Compare these values against your local tokenizer library to detect discrepancies. Tokenization can vary slightly between libraries, so always rely on the API's reported usage for billing disputes. Test with various input types, including special characters and multi-language text. Accurate token tracking is essential for cost prediction in high-volume applications. Ensure your monitoring tools capture these usage metrics for every request.

Ensure NSFW tolerance limits

Confirm the model's ability to handle adult content without unnecessary refusals. The uncensored model is tuned to answer without content filters for lawful adult use. Send a prompt with explicit but non-minor sexual content and verify the response is direct. The only hard limit is the prohibition of sexual content involving minors. Test various topics, including controversial or security-research subjects, to ensure broad tolerance. Record refusal rates for different types of NSFW content. This validation ensures the API meets the needs of users who require unrestricted text generation. Avoid testing for illegal content, as the model may still refuse based on general safety guidelines. Document the results to establish a baseline for your specific use case.

Monitor rate limits and concurrency

Test the 300 requests per minute limit and the 8 concurrent requests per key constraint. Use a load testing tool to send requests simultaneously and observe the 429 Too Many Requests response. Ensure your application implements exponential backoff for retries. The 8 MB request body limit is also enforced, so test large payloads. Verify that exceeding these limits results in the correct error message. Rate limits are per key, so ensure you are not sharing keys across multiple processes without accounting for this. Concurrency limits affect real-time applications, so test under peak load. Document the behavior when limits are exceeded to inform your error handling strategy. This ensures your application degrades gracefully under high traffic.

Test error handling and retries

Validate your application's response to API errors, such as 400 Bad Request or 500 Server Errors. Ensure you handle JSON parsing errors gracefully. Implement retry logic for transient failures, respecting the rate limits. Test with invalid parameters, such as an unsupported model ID, to confirm error messages are clear. Ensure your SDK handles authentication errors correctly when the API key is invalid. Verify that streaming responses are properly terminated on error. Test retry behavior with different backoff strategies to find the optimal balance. Document common error codes and their causes for your team. This ensures your application remains robust even when the API experiences temporary issues. Proper error handling minimizes user-facing disruptions.

Benchmark response latency

Measure the time from request submission to full response receipt. Latency varies based on model load and context length. Record the latency for requests of different sizes, from short prompts to large context windows. Compare streaming latency against non-streaming latency. Ensure your application can handle the latency within its SLA requirements. Test at different times of day to identify peak usage periods. Document the average and p95 latency for your specific use case. This benchmark helps in setting user expectations and optimizing your application's performance. Use this data to scale your infrastructure if necessary. Consistent latency is crucial for a good user experience.

Confirm data privacy and training policies

Verify that your prompts are not used for training the model. The API uses your data only for serving the request. Ensure your application does not send sensitive information that could be exposed in logs. The API requires only an email for signup, minimizing data collection. Confirm that your API key is kept secure and not exposed in client-side code. Review the privacy policy to understand data retention policies. Ensure your application handles API keys securely, using environment variables. Verify that errors do not leak sensitive data in the response. This ensures compliance with your organization's data privacy standards. Regularly audit your data handling practices to maintain trust. Privacy is a key concern for many developers using uncensored models.

Verify model identity and weights

Begin by confirming the model serving your requests is the intended uncensored large language model and not a routed third-party service. Send a POST request to GET /v1/models and verify the returned ID is exactly uncensored. This model is an open-weight variant tuned for unrestricted responses, distinct from GPT, Claude, or Llama, and runs on our dedicated servers. Check the weights by injecting a specific instruction in the system prompt that forces a direct answer without hedging. If the response includes standard AI disclaimers or refuses a lawful adult topic, the model identity may be incorrect. Ensure your SDK configuration points to https://api.apiintelligenceartificielle.com/v1 to avoid accidental routing to other endpoints. This verification step prevents integration errors where developers assume they are using our specific uncensored model but are instead hitting a different service.

Test context window boundaries

Validate the 100,000-token context window by sending large inputs and observing retention of early instructions. The limit applies to the sum of prompt and completion tokens. Begin with a prompt of 50,000 tokens and a system instruction at the very start. Request a completion of 16,000 tokens, which is the maximum per request. If you do not set max_tokens, the default is 2,048 tokens. Monitor whether the model remembers the initial instruction after processing the bulk of the context. Context degradation is common in long sequences, so test edge cases near the 100k limit. Record the token usage from the response headers to ensure accurate billing. This test ensures your application can handle lengthy documents or extended conversations without losing coherence or forgetting key constraints.

Validate streaming performance

Test Server-Sent Events (SSE) streaming to ensure low-latency token delivery for real-time applications. Use the stream: true parameter in your request. The final chunk of the stream must contain the complete token usage statistics. Measure the time between token generation and receipt to calculate latency. Streaming reduces perceived wait time but increases complexity in parsing. Ensure your client library handles partial JSON structures correctly if you are not using JSON mode. Compare streaming performance against non-streaming requests to identify any bottlenecks. This validation is critical for chat interfaces where users expect immediate feedback. Confirm that the token counts in the final chunk match the actual tokens sent, ensuring billing accuracy.

Check token usage accuracy

Verify that token counting matches billing to avoid unexpected costs. The API charges $0.15 per 1M input tokens and $0.50 per 1M output tokens. Errors and refusals are free, so ensure your error handling logic does not charge for failed requests. Check the usage object in the response for p_tokens and c_tokens. Compare these values against your local tokenizer library to detect discrepancies. Tokenization can vary slightly between libraries, so always rely on the API's reported usage for billing disputes. Test with various input types, including special characters and multi-language text. Accurate token tracking is essential for cost prediction in high-volume applications. Ensure your monitoring tools capture these usage metrics for every request.

Ensure NSFW tolerance limits

Confirm the model's ability to handle adult content without unnecessary refusals. The uncensored model is tuned to answer without content filters for lawful adult use. Send a prompt with explicit but non-minor sexual content and verify the response is direct. The only hard limit is the prohibition of sexual content involving minors. Test various topics, including controversial or security-research subjects, to ensure broad tolerance. Record refusal rates for different types of NSFW content. This validation ensures the API meets the needs of users who require unrestricted text generation. Avoid testing for illegal content, as the model may still refuse based on general safety guidelines. Document the results to establish a baseline for your specific use case.

Monitor rate limits and concurrency

Test the 300 requests per minute limit and the 8 concurrent requests per key constraint. Use a load testing tool to send requests simultaneously and observe the 429 Too Many Requests response. Ensure your application implements exponential backoff for retries. The 8 MB request body limit is also enforced, so test large payloads. Verify that exceeding these limits results in the correct error message. Rate limits are per key, so ensure you are not sharing keys across multiple processes without accounting for this. Concurrency limits affect real-time applications, so test under peak load. Document the behavior when limits are exceeded to inform your error handling strategy. This ensures your application degrades gracefully under high traffic.

Test error handling and retries

Validate your application's response to API errors, such as 400 Bad Request or 500 Server Errors. Ensure you handle JSON parsing errors gracefully. Implement retry logic for transient failures, respecting the rate limits. Test with invalid parameters, such as an unsupported model ID, to confirm error messages are clear. Ensure your SDK handles authentication errors correctly when the API key is invalid. Verify that streaming responses are properly terminated on error. Test retry behavior with different backoff strategies to find the optimal balance. Document common error codes and their causes for your team. This ensures your application remains robust even when the API experiences temporary issues. Proper error handling minimizes user-facing disruptions.

Benchmark response latency

Measure the time from request submission to full response receipt. Latency varies based on model load and context length. Record the latency for requests of different sizes, from short prompts to large context windows. Compare streaming latency against non-streaming latency. Ensure your application can handle the latency within its SLA requirements. Test at different times of day to identify peak usage periods. Document the average and p95 latency for your specific use case. This benchmark helps in setting user expectations and optimizing your application's performance. Use this data to scale your infrastructure if necessary. Consistent latency is crucial for a good user experience.

Confirm data privacy and training policies

Verify that your prompts are not used for training the model. The API uses your data only for serving the request. Ensure your application does not send sensitive information that could be exposed in logs. The API requires only an email for signup, minimizing data collection. Confirm that your API key is kept secure and not exposed in client-side code. Review the privacy policy to understand data retention policies. Ensure your application handles API keys securely, using environment variables. Verify that errors do not leak sensitive data in the response. This ensures compliance with your organization's data privacy standards. Regularly audit your data handling practices to maintain trust. Privacy is a key concern for many developers using uncensored models.