[FEATURE] Model class for Bedrock Custom Model Import (requires Invoke not Converse)

[austinmw]

Problem Statement

Bedrock Custom Model import uses Invoke API not Converse API. Currently no model classes are compatible with, e.g., deploying a Qwen3 model to Bedrock Custom Model Import and using this model with Strands.

Proposed Solution

Add a model class for this

Use Case

Bedrock Custom Model Import

Alternatives Solutions

No response

Additional Context

No response

[iamgautamraj]

Hey there!

I'm interested in contributing to this feature. I've been studying the codebase and understand that:

1. The current BedrockModel uses the Converse API which doesn't support Custom Model Import
2. Custom imported models (like Qwen3, Llama, Mistral) require the Invoke API (invoke_model_with_response_stream)
3. This would need a new model class with adapter patterns since each imported model has different input/output formats

Before I start working on a design proposal, I have a few clarifying questions:

1. Priority Models

Which custom model formats should be prioritized for the initial implementation?

• Qwen3
• Llama 3.x
• Mistral
• Or should I target a generic implementation??

2. Tool Calling Scope

Since the Invoke API doesn't have native tool calling support like Converse, should tool calling be:

• In-scope for v1 (handled via prompt engineering/parsing)?
• Out-of-scope for initial implementation?

3. Adapter Pattern

I'm thinking of a flexible design where users can provide custom input/output formatters for their specific model. Something like:

from strands.models import BedrockCustomModel

model = BedrockCustomModel(
    model_arn="arn:aws:bedrock:...",  # Custom imported model ARN
    model_format="llama3",            # Or provide custom formatter
)

Would this approach align with the project's design philosophy?

4. Existing Work

Are there any existing designs, PRs, etc. on this that I should review before proposing a solution?

If this approach sounds good, please assign this issue to me, and I can start working on the initial PR!