MiniMax M2, M2.1, and M2.5 now available on Amazon Bedrock

Amazon Bedrock has continued expanding its model catalog: MiniMax M2, M2.1, and M2.5 have joined the platform — three open-weight models from Chinese developer MiniMax, each optimized for a different task profile. The integration brings support for tool-calling, implicit prompt caching, and multi-turn agentic workflows, with models immediately available in 14 AWS regions without a waitlist or special approvals.

Three models, three focuses

Each of the three models offers different trade-offs between context capacity, architecture, and intended use.

MiniMax M2 offers a record context window of 1 million tokens — sufficient for processing complete business documentation, long legal contracts, or months of conversation history in a single call. The maximum output is 8,192 tokens, consistent across all three variants. M2 is a natural choice for tasks requiring long-term contextual consistency: corporate archive analysis, summarization of extensive research papers, or tracking long software repositories where complete context is critical for accurate conclusions.

MiniMax M2.1 reduces the window to 196,384 tokens, but compensates with improved reasoning depth and precision on coding tasks. It is trained for reliable execution of complex multi-step instructions, excelling at mathematical proofs, formal code verification, and analysis that must be accurate at the detail level. This variant is the preferred choice when logical consistency matters more than raw context scope.

MiniMax M2.5 introduces an architectural innovation: Mixture-of-Experts (MoE) with a total of 230 billion parameters, of which only 10 billion are active per token. The model carries the knowledge and capacity of a 230B network, at the computational cost of a dense 10B-parameter network — making it more economical than comparable dense models of the same depth.

How is the M2.5 architecture built?

The Mixture-of-Experts approach divides the model into a set of specialized sub-networks (experts). For each input token, a gating network selects a small subset of experts to process it — in M2.5’s case, that subset corresponds to an active capacity of 10 billion parameters, while the rest of the network remains inactive for that specific token but is available for others in the same sequence.

What particularly distinguishes M2.5 is the reinforcement learning phase focused on agentic execution logic: the model was trained directly on scenarios involving tool calls, error handling, branching, and timeout management. These are precisely the edge cases where smaller models and models trained exclusively on text fail — M2.5 treats them as a first-class training signal, making it more robust for autonomous multi-step workflows.

Service tiers and integration

Amazon Bedrock offers three access tiers for all MiniMax models:

• Priority — up to 25% better latency compared to Standard; designed for real-time systems and mission-critical applications
• Standard — the default on-demand tier for balanced cost and performance
• Flex — reduced pricing for batch and latency-tolerant tasks such as overnight batch jobs and bulk processing

Two API interfaces are available for integration. The bedrock-mantle endpoint offers a Chat Completions format compatible with the OpenAI SDK — organizations with existing OpenAI integrations can switch the URL and model ID without changing any other code. The bedrock-runtime endpoint offers the native AWS interface: Converse API and InvokeModel, with full AWS SDK support in Python, TypeScript, Java, and other languages.

Implicit prompt caching is active without special configuration: repeated system prompt or document segments are automatically cached on the platform side, reducing latency and costs in iterative conversations and agentic loops. Tool-calling and function invocation are available across all three models using standard JSON schema syntax for tool definitions.

Regional availability

Models are available in 14 AWS regions: US East (N. Virginia, Ohio), US West (Oregon), European regions (Frankfurt, Stockholm, Milan, Ireland, London), Asia-Pacific regions (Tokyo, Mumbai, Sydney, Jakarta, Melbourne), and South America (São Paulo). This distribution covers the key regulatory zones requiring data residency within specific geographic boundaries — particularly relevant for GDPR-regulated European workloads and Asian markets with local data storage requirements.

MiniMax M2, M2.1, and M2.5 are available immediately on Amazon Bedrock — with no special prerequisites beyond a standard AWS account with appropriate permissions.