AWS launches MiniMax M2 family on Amazon Bedrock with 1M token context and MoE architecture
Amazon Web Services has added three MiniMax models to Amazon Bedrock: M2, M2.1, and M2.5. The newest model, M2.5, uses a mixture-of-experts architecture with 230 billion total parameters and 10 billion active per token, trained specifically for agent-native execution and coding tasks.
MiniMax-M2.5 at a glance
How MiniMax-M2.5 ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| Arena Elo | 1391 Elo | 110 of 146 | LMArena, Nov 2026 ↗ |
| GPQA | 84.8% | 91 of 165 | Artificial Analysis, Jul 2026 ↗ |
| SWE-bench Verified | 80.2% | 13 of 80 | llm-stats, Jul 2026 ↗ |
| Speed (tok/s) | 49 tok/s | 91 of 105 | Artificial Analysis, Nov 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
MiniMax M2 Family Now Available on Amazon Bedrock
Amazon Web Services has integrated three models from MiniMax's M2 family into Amazon Bedrock, giving developers access to open-weight models optimized for software engineering and agentic use cases. The models run entirely on AWS infrastructure with data isolation guarantees.
Model Specifications
The three models differ in context windows and training focus:
MiniMax M2 (minimax.minimax-m2): 1 million token context window, 8K max output tokens, trained for multilingual text generation, reasoning, and coding.
MiniMax M2.1 (minimax.minimax-m2.1): 196K token context window, 8K max output tokens, with improvements to reasoning depth, coding accuracy, and instruction following.
MiniMax M2.5 (minimax.minimax-m2.5): 196K token context window, 8K max output tokens, 230 billion total parameters with 10 billion active per token. According to AWS, M2.5 was trained specifically for agent-native execution using reinforcement learning on agentic scaffolds.
All three models use a mixture-of-experts (MoE) architecture where only a fraction of parameters activate per token. For M2.5, this means the model delivers the capacity of a 230B parameter model while consuming compute equivalent to only 10B parameters per forward pass.
Pricing and Service Tiers
Pricing per million tokens was not disclosed in the announcement. AWS offers three service tiers for all models: Standard, Priority, and Flex. According to AWS, on-demand inference automatically scales to handle workload requirements.
Access Methods
AWS provides two endpoints for accessing the models:
bedrock-mantle endpoint (recommended): Uses the Chat Completions API, compatible with OpenAI Python and TypeScript SDK interfaces. Supports Amazon Bedrock API keys, projects, and client-side tool calling.
bedrock-runtime endpoint: Uses Converse and InvokeModel APIs via the AWS SDK. Required for native Bedrock features including Guardrails, Agents, Flows, and model evaluation.
The models are available immediately in the Amazon Bedrock console playground and via both API endpoints.
Data Protection
AWS states that prompts and completions are not used to train any models and content is not shared with MiniMax. Inference runs entirely on AWS-operated infrastructure.
What This Means
The addition of MiniMax models expands Bedrock's open-weight options beyond Meta's Llama and Mistral families. The 1 million token context window in M2 positions it for document analysis workloads, while M2.5's agent-focused training addresses the growing demand for tool-calling models. The MoE architecture's 10B active parameters per token could provide cost advantages over dense models of similar capability, though without public pricing comparisons are speculative. The dual endpoint system reflects AWS's evolution toward OpenAI-compatible APIs while maintaining backward compatibility with existing Bedrock integrations.
Related Articles
Kimi Code CLI v2.7.0 adds native Amazon Bedrock provider with standard AWS credentials
Kimi Code CLI v2.7.0, Moonshot AI's open source coding agent, adds a native bedrock provider. It resolves the standard AWS credential chain, including named profiles and SSO, so users no longer need a separate Amazon Bedrock API key. AWS demonstrates the setup with Kimi K3.1 Pro Highspeed.
AWS launches SageMaker JumpStart Model Leaderboard covering more than 70 models
Amazon Web Services introduced the Amazon SageMaker JumpStart Model Leaderboard on November 18, 2026. It gives SageMaker Studio users one view of benchmark results, response performance and estimated instance pricing across more than 70 models.
Amazon Polly voice cloning is generally available as a gated release in seven languages
Amazon Polly voice cloning is now generally available as a gated release. It builds a custom voice on the Polly generative engine from a few minutes of one consenting speaker's audio, in seven languages. Building a voice is free, and usage is billed pay-as-you-go.
Cline v4.1.27 fixes reasoning-effort errors that caused GPT-6 and gpt-5.6 tool-call rejections
Cline v4.1.27 fixes several reasoning-effort bugs that caused GPT-6 and gpt-5.6 endpoints to reject tool calls. It also refreshes the model catalog with 94 new providers and changes default models for 67 others.
Comments
Loading...