Updated model support: You can now import Qwen3 models with model import.

Amazon Bedrock Custom Model Import

You can now import customized open-source models into Amazon Bedrock using the Custom Model Import feature. This allows you to leverage Amazon Bedrock features to make inference calls to your customized models.

Supported Regions

• US East (N. Virginia)
• US West (Oregon)
• Europe (Frankfurt)

Limitations

• Not supported with: Batch inference, AWS CloudFormation

Supported Model Patterns

• Fine-tuned or Continued Pre-training model
• Adaptation
• Pretrained from scratch

Supported Architectures

• Mistral
• Mixtral
• Flan
• Llama 2, Llama3, Llama3.1, Llama3.2, Llama 3.3, and Mllama
• GPTBigCode
• Qwen2, Qwen2.5, Qwen2-VL, Qwen2.5-VL, Qwen3

Model Import Requirements

• Model Weights: Less than 100GB for multimodal models and 200GB for text models
• Maximum Context Length: Less than 128K
• Transformer Version: 4.51.3

Importing a Model from Amazon S3

To import a model, create a model import job in the Amazon Bedrock console or API, specifying the Amazon S3 URI for the model files. The import job will automatically detect the model's architecture.

Required Model Files

• .safetensors - Model weights in Safetensor format
• config.json
• tokenizer_config.json
• tokenizer.json
• tokenizer.model

Supported Tokenizers

• T5Tokenizer
• T5TokenizerFast
• LlamaTokenizer
• LlamaTokenizerFast
• CodeLlamaTokenizer
• CodeLlamaTokenizerFast
• GPT2Tokenizer
• GPT2TokenizerFast
• GPTNeoXTokenizer
• GPTNeoXTokenizerFast
• PreTrainedTokenizer
• PreTrainedTokenizerFast
• Qwen2Tokenizer
• Qwen2TokenizerFast

What to do

• Ensure your model complies with applicable terms and licenses.
• Check the Amazon Bedrock pricing for custom model import costs.

Source: AWS release notes