MiniMax M2.5 on Amazon Bedrock: models, APIs and limits

Amazon Bedrock offers the MiniMax M2 family as fully managed, open-weight foundation models for production use, with three models now available: MiniMax M2, MiniMax M2.1 and MiniMax M2.5. The service runs inference entirely on AWS-operated infrastructure, and "your prompts and completions are not used to train any models, and your content is not shared with the model providers."

What models are available and how do they differ?

MiniMax on Bedrock includes three M2-series models; their core differences are context window, training focus and suitability for workloads. MiniMax M2 (minimax.minimax-m2) offers a 1M token context window and targets long-context or multilingual general-purpose tasks. MiniMax M2.1 (minimax.minimax-m2.1) and MiniMax M2.5 (minimax.minimax-m2.5) each provide a 196K token context window and focus on deeper reasoning and agent-native execution respectively. All three list a maximum output of 8K tokens and the same service tiers: Standard, Priority, Flex.

MiniMax M2.5 is described as purpose-built for agent-native execution, trained with reinforcement learning on agentic scaffolds and optimized for tool-calling, multi-step task decomposition and long-horizon coding tasks. M2.5 uses a mixture-of-experts architecture with "230 billion total parameters and 10 billion active per token," delivering large-model knowledge capacity while routing compute proportional to the 10B active parameters per forward pass.

How do you access MiniMax models on Amazon Bedrock and what APIs are available?

You can invoke MiniMax models through two Bedrock endpoints: bedrock-mantle and bedrock-runtime, each serving different integration needs. The bedrock-mantle endpoint (https://bedrock-mantle.{region}.api.aws/v1) is the public API for Bedrock's next-generation inference engine; it uses the Chat Completions API and is recommended for most workloads. It supports Amazon Bedrock API keys, projects, and client-side tool calling and is compatible with the OpenAI Python and TypeScript SDKs by updating base URL and model ID.

The bedrock-runtime endpoint (https://bedrock-runtime.{region}.amazonaws.com) exposes the Converse and InvokeModel APIs via the AWS SDK. Use bedrock-runtime for native Bedrock features currently available there, including Guardrails, Agents, Flows, and model evaluation. Example usage in the documentation demonstrates calling minimax.minimax-m2.5 with the OpenAI Python SDK pointed at the bedrock-mantle base URL and retrieving an Amazon Bedrock API key from AWS Secrets Manager. For production, the documentation recommends short-term API keys that expire automatically (maximum 12 hours) and inherit the permissions of the IAM role that generated them.

The docs also show a tool-calling workflow: define functions the model can invoke, send tool definitions with the user message, inspect the assistant response for tool calls, run the corresponding function in your application, then pass the function result back to the model so it can produce a final natural-language answer.

Why it matters

Amazon Bedrock combines access to open-weight MiniMax models with AWS operational and data-protection guarantees. Customers get the ability to evaluate model architecture and training methodology, fine-tune on proprietary data, and run inference on AWS infrastructure without sharing prompts or completions with external model providers. For organizations that require control over data flows and compliance, Bedrock's endpoints and short-term API key model provide concrete operational controls while exposing models designed specifically for agentic workflows and long-context tasks.

What to watch

Watch for updates to the Bedrock MiniMax catalog in the Amazon Bedrock model documentation as new MiniMax versions become available, and monitor whether additional Bedrock-native features move from bedrock-runtime to bedrock-mantle. Also track pricing and per-token charges on the Amazon Bedrock pricing page, since each model invocation incurs per-token costs.