AWS bedrock documentation change
Summary
Updated DeepSeek-R1 documentation to add guidance about limiting max_tokens to 8,192 for optimal quality (while maintaining API support up to 32,768) and fixed URL formatting
Security assessment
The changes focus on improving response quality recommendations and parameter documentation. While the max_tokens limitation could indirectly prevent resource overuse, there's no explicit mention of security vulnerabilities, attacks, or access control changes. The IAM policy reference remains unchanged except for URL formatting.
Diff
diff --git a/bedrock/latest/userguide/model-parameters-deepseek.md b/bedrock/latest/userguide/model-parameters-deepseek.md index 995fd2348..1d70bbbf9 100644 --- a//bedrock/latest/userguide/model-parameters-deepseek.md +++ b//bedrock/latest/userguide/model-parameters-deepseek.md @@ -13 +13,6 @@ When you make inference calls with DeepSeek’s models, you must include a promp -You can't remove request access from the Amazon Titan, Amazon Nova, DeepSeek-R1, Mistral AI, Meta Llama 3 Instruct, and Meta Llama 4 models. You can prevent users from making inference calls to these models by using an IAM policy and specifying the model ID. For more information, see [Deny access for inference of foundation models](https://docs.aws.amazon.com/bedrock/latest/userguide/security_iam_id-based-policy-examples.html#security_iam_id-based-policy-examples-deny-inference .html). + * You can't remove request access from the Amazon Titan, Amazon Nova, DeepSeek-R1, Mistral AI, Meta Llama 3 Instruct, and Meta Llama 4 models. You can prevent users from making inference calls to these models by using an IAM policy and specifying the model ID. For more information, see [Deny access for inference of foundation models](https://docs.aws.amazon.com/bedrock/latest/userguide/security_iam_id-based-policy-examples.html#security_iam_id-based-policy-examples-deny-inference .html). + + * For optimal response quality with DeepSeek-R1, limit the `max_tokens` parameter to 8,192 tokens or fewer. While the API accepts up to 32,768 tokens, response quality significantly degrades above 8,192 tokens. This aligns with the model's reasoning capabilities as described in the [inference reasoning guide](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-reasoning.html). + + + @@ -51 +56 @@ DeepSeek has the following inference parameters for a Text Completion inference - * **max_tokens** – (int) Tokens used, minimum of 1 to a max of 32,768 tokens. + * **max_tokens** – (int) Tokens used, minimum of 1 to a max of 8,192 tokens for optimal quality. While the API accepts up to 32,768 tokens, response quality significantly degrades above 8,192 tokens.