AWS BedRock + DeepSeek = very poor experience

What you're experiencing with DeepSeek on AWS Bedrock is not normal behavior. The random characters in the response indicate there might be an issue with how you're interacting with the model.

DeepSeek R1 is available on Amazon Bedrock, but there are a few things to consider:

1. Make sure you're using the correct model ARN when making API calls. The proper ARN format for DeepSeek would be something like "arn:aws:bedrock:region::foundation-model/deepseek.r1-v1:0".

2. Check that you're using the appropriate request format for the model. Different models may have specific input formatting requirements.

3. If you're using the RetrieveAndGenerate API with DeepSeek and a Knowledge Base, ensure your configuration is correct. DeepSeek does support Knowledge Base integration with Amazon Bedrock.

4. Consider trying a different model temporarily to see if the issue is specific to DeepSeek or more general.

The garbled output you're seeing suggests there might be an encoding issue or a problem with how the request is being processed. I'd recommend checking your API call format and parameters, or reaching out to AWS Support if the problem persists.
Sources
Deepseek RetrieveandGenerate | AWS re:Post
Bedrock and Deepseek-R1 | AWS re:Post
Working with models not yet optimized for Amazon Bedrock Agents - Amazon Bedrock

Not normal. Can you retry it? Also what parameters are you using?

I retried just now, with default params: same result.

Region: EU - Stockholm System prompt = <blank> Model reasoning = off Maximum output tokens = 4096 Stop sequences = <blank> Temperature = 1 Top P = 1 Prompt caching = off

Today I received an alert: "Between November 22 11:00 PM and November 26 6:48 AM PST customers using the DeepSeek V3.1 model experienced timeouts and malformed responses in the EU-NORTH-1 Region. Engineers have determined the cause as a latent bug in the vLLM engine that caused the model to degrade under load. A mitigation was applied to resolve impact. No further customer action is required. The issue has been resolved and the service is operating normally."

Tested, seems to be ok.

Thank you

I am also experiencing degraded under heavy load in us-east-2 region (default params, flex). Do you have any knowledge if the problem is also happening in other regions?

Thank you,