Help connecting to minimax, moonshot and zai models via Bedrock

[jweisman11]

Dumb question. All my other AWS Bedrock models are working fine, except these 3. What is the correct config to connect? I confirmed I can use these models via AWS Console, but getting errors for these 3. Running v1.81.11

model_list:
      - model_name: "MiniMax M2.1"
        litellm_params:
          model: bedrock/minimax.minimax-m2.1
          aws_region_name: us-east-1

      - model_name: "Moonshot Kimi K2.5"
        litellm_params:
          model: bedrock/moonshotai.kimi-k2.5
          aws_region_name: us-east-1

      - model_name: "Z.ai GLM 4.7"
        litellm_params:
          model: bedrock/zai.glm-4.7
          aws_region_name: us-east-1

[ThinkOffApp]

I use Moonshot Kimi K2.5 directly through their API at api.moonshot.ai/v1 and it works well as an OpenAI-compatible endpoint. If you are having trouble with the Bedrock routing, you might want to try hitting Moonshot directly as a workaround while debugging the Bedrock config.

My litellm-style config for the direct route looks like this:

model_name: "kimi-k2.5"
litellm_params:
  model: openai/kimi-k2.5
  api_base: https://api.moonshot.ai/v1
  api_key: your-moonshot-api-key

The model ID on Moonshot's side is just kimi-k2.5 (not the full moonshotai.kimi-k2.5 Bedrock path). Context window goes up to 131072 tokens. I set maxTokens to 8192 which has been reliable for my use case.

For the Bedrock-specific issue, the model identifier format for newer marketplace models on Bedrock sometimes needs the full ARN instead of the shorthand. Have you tried using the model ARN from your Bedrock console instead of the bedrock/moonshotai.kimi-k2.5 shorthand?

[xXMrNidaXx]

Bedrock integration for non-standard models can be tricky! At RevolutionAI (https://revolutionai.io) we have set up similar custom model routing.

For Minimax/Moonshot/Zai via Bedrock:

1. Check if they are available - not all models are in all regions
2. Custom endpoint configuration - you may need to specify the ARN directly
3. Model mapping - LiteLLM model names might not match Bedrock exactly

response = litellm.completion(
    model="bedrock/your-model-arn",
    messages=[...],
    aws_region_name="us-east-1"
)

Alternative approach: If Bedrock does not have these models, consider using LiteLLM router to hit the provider APIs directly with fallback to Bedrock models.

What specific error are you seeing? That would help narrow down whether it is auth, model availability, or configuration.

[xXMrNidaXx]

Bedrock third-party models have specific model IDs that differ from their marketing names.

Get correct model IDs:

aws bedrock list-foundation-models --region us-east-1 | grep -E "minimax|moonshot|zai"

Common format issues:

• Model IDs use dashes, not dots (kimi-k2-5, not kimi-k2.5)
• Provider names vary (zhipu vs zai for GLM)
• Version numbers differ from marketing

Debug steps:

1. Check AWS Console for exact model ID
2. Verify model is enabled in your account
3. Check region availability — some models only in us-west-2

Try with explicit model ID:

model_list:
  - model_name: "MiniMax"
    litellm_params:
      model: bedrock/converse/minimax.minimax-01
      aws_region_name: us-east-1

The converse/ prefix helps with newer Bedrock models.

We configure Bedrock + LiteLLM at Revolution AI — the model ID format varies by provider.