Hello.

I ran the following command in my AWS account and confirmed a normal response.
This suggests a possible issue, such as a temporary shortage of Bedrock capacity on the AWS side.
As of 2026/04/23 01:57 PM, it was executed successfully.

~ $ aws bedrock-runtime converse \
>   --model-id moonshotai.kimi-k2.5 \
>   --messages '[{"role":"user","content":[{"text":"Return valid JSON in the format {\"word\": \"HELLO\"}. Nothing else."}]}]' \
>   --inference-config '{"maxTokens":200,"temperature":0.1}' \
>   --region us-east-1
{
    "output": {
        "message": {
            "role": "assistant",
            "content": [
                {
                    "text": " {\"word\": \"HELLO\"}"
                }
            ]
        }
    },
    "stopReason": "end_turn",
    "usage": {
        "inputTokens": 44,
        "outputTokens": 8,
        "totalTokens": 52
    },
    "metrics": {
        "latencyMs": 596
    }
}
~ $ aws bedrock-runtime converse \
>   --model-id moonshot.kimi-k2-thinking \
>   --messages '[{"role":"user","content":[{"text":"Return valid JSON in the format {\"word\": \"HELLO\"}. Nothing else."}]}]' \
>   --inference-config '{"maxTokens":200,"temperature":0.1}' \
>   --region us-east-1
{
    "output": {
        "message": {
            "role": "assistant",
            "content": [
                {
                    "reasoningContent": {
                        "reasoningText": {
                            "text": " The user wants me to return valid JSON in a specific format: {\"word\": \"HELLO\"}.\nThey explicitly say \"Nothing else.\"\n\nSo I should just output that exact JSON string and nothing else. No extra text, no explanations, no markdown formatting.\n\nThe JSON is simple and valid. I will just return it as is."
                        }
                    }
                },
                {
                    "text": " {\"word\": \"HELLO\"}"
                }
            ]
        }
    },
    "stopReason": "end_turn",
    "usage": {
        "inputTokens": 24,
        "outputTokens": 77,
        "totalTokens": 101
    },
    "metrics": {
        "latencyMs": 1468
    }
}

Based on the information available about Amazon Bedrock's supported models, I can confirm that both Moonshot AI models you're referencing are indeed supported in Amazon Bedrock:

Kimi K2.5 (Model ID: moonshotai.kimi-k2.5) is Moonshot AI's multimodal model with improved reasoning, coding, and multilingual capabilities. It supports text and image inputs, text outputs, and streaming, and is available in us-east-1 among other regions.

Kimi K2 Thinking (Model ID: moonshot.kimi-k2-thinking) is Moonshot AI's reasoning model with chain-of-thought capabilities designed for complex problem solving in math, coding, and logic. It supports text input and output with streaming capabilities, and is also available in us-east-1.

However, the specific runtime behavior you're describing—where the models are returning only padding characters (!) and exhausting token budgets—represents an operational issue that is not documented in the available model specifications. The models are listed as having "N/A" for inference parameters, which means specific parameter controls may be limited or handled differently than other model families.

Given that you've documented a working state earlier the same day followed by this regression, and you've systematically ruled out configuration issues through your testing, this appears to be a service-side issue that requires investigation by AWS Support. I recommend opening a support case with AWS, providing the detailed reproduction steps, timestamps of the working vs. non-working states, and the specific model invocation logs you've collected. This will allow the Bedrock service team to investigate potential model deployment or infrastructure issues in the us-east-1 region.
Sources
Supported foundation models in Amazon Bedrock - Amazon Bedrock
Moonshot AI - Amazon Bedrock

Running into the exact same problem with Kimini K2.5. First it was only occurring sporadically on larger conversations. Often retrying it could resolve the issue, with the disadvantage of additional input costs. For around 24 hours, Kimini K2.5 has been completely unusable for me, as it affects now every conversation, no matter how small or how many tries. Could even reproduce it in the playground: