What are the key differences between MiniMax M2, M2.1, and M2.5?

MiniMax M2 has a 1 million token context window and is best for long-context or multilingual general-purpose work. M2.1 has a 196K token context window and improved reasoning, coding, and instruction following. M2.5 also has 196K tokens but is trained specifically for agent-native execution, tool-calling, and coding-heavy workloads. All three support 8K max output tokens.

How does the mixture-of-experts architecture reduce inference cost?

MiniMax M2.5 has 230 billion total parameters but only 10 billion activate per token. The mixture-of-experts routing mechanism provides the knowledge capacity of a 230B model while consuming compute proportional to only 10B active parameters per forward pass, significantly reducing inference cost.

Is customer data shared with MiniMax or used for model training?

No. Amazon Bedrock's fully managed service runs inference entirely on AWS-operated infrastructure. Prompts and completions are not used to train any models, and content is not shared with model providers.

Back to articlesLarge Language Models

Large Language Models

MiniMax models now available on Amazon Bedrock

Amazon AI Blog2h ago7 min read

Key takeaway

Amazon Bedrock now offers three open-weight MiniMax AI models, with the newest M2.5 specifically designed for agent-native execution and autonomous tool-calling. The models run entirely on AWS infrastructure, ensuring customer data is not shared with model providers or used for training. Organizations can access them through two API endpoints optimized for different use cases—one compatible with OpenAI SDKs and one for native AWS features.

Summaries like this, in your inbox every morning.

3 Key Points

What happened
Amazon Bedrock added three MiniMax open-weight AI models to its fully managed service. The newest, MiniMax M2.5, is trained specifically for agent-native execution (AI that makes decisions and calls tools autonomously). All three models run on AWS infrastructure, and customer data is not shared with model providers or used for training.
Why it matters
Organizations deploying AI workloads to production need models that combine strong capabilities with data protection and compliance guarantees. These MiniMax models use a mixture-of-experts architecture, meaning they deliver the capability of a much larger model while consuming compute proportional to only 10 billion active parameters per forward pass—reducing inference cost. They are purpose-built for software engineering, agentic workflows, and long-context document analysis, so businesses in those domains can now access them without managing infrastructure.
What to watch
Two endpoints provide access—bedrock-mantle (recommended, compatible with OpenAI SDK) and bedrock-runtime (for native AWS features like Guardrails and Agents). All three service tiers (Standard, Priority, Flex) support each model. MiniMax M2 offers 1 million token context window; M2.1 and M2.5 each support 196K tokens and 8K max output tokens.

Context & Analysis

Amazon Bedrock's addition of MiniMax models reflects a broader shift toward giving enterprises access to frontier open-weight foundation models without sacrificing data protection or operational control. Organizations experimenting with AI have increasingly needed production-ready environments that guarantee security and regulatory compliance—requirements that drive the choice between model hosting platforms. By offering three models from the same family, Bedrock lets customers match workload demands (long-context multilingual tasks, complex reasoning, or agentic tool-calling) to the right model, rather than forcing a one-size-fits-all choice.

The mixture-of-experts design is a key practical advantage. A 230B-parameter model that activates only 10B parameters per token delivers capacity at a fraction of the compute cost, making expensive production inference more economical for developers and enterprises. MiniMax's focus on software engineering and agentic execution aligns with market demand: customers building coding assistants and autonomous workflows can now use models trained specifically for those tasks through a managed AWS service, eliminating infrastructure provisioning and model hosting overhead. The dual endpoint architecture (bedrock-mantle for OpenAI SDK compatibility and bedrock-runtime for AWS-native features like Guardrails) gives teams flexibility in how they integrate these models into existing workflows.

FAQ

What are the key differences between MiniMax M2, M2.1, and M2.5?: MiniMax M2 has a 1 million token context window and is best for long-context or multilingual general-purpose work. M2.1 has a 196K token context window and improved reasoning, coding, and instruction following. M2.5 also has 196K tokens but is trained specifically for agent-native execution, tool-calling, and coding-heavy workloads. All three support 8K max output tokens.
How does the mixture-of-experts architecture reduce inference cost?: MiniMax M2.5 has 230 billion total parameters but only 10 billion activate per token. The mixture-of-experts routing mechanism provides the knowledge capacity of a 230B model while consuming compute proportional to only 10B active parameters per forward pass, significantly reducing inference cost.
Is customer data shared with MiniMax or used for model training?: No. Amazon Bedrock's fully managed service runs inference entirely on AWS-operated infrastructure. Prompts and completions are not used to train any models, and content is not shared with model providers.

Discussion

No comments yet. Be the first to share your thoughts!

S&P Global Restructures Market Intelligence Unit Into Two Verticals to Boost AI Capabilities

Top Companies AI — US (2/2)20m ago

Economist maps AI worldviews via values survey; enterprise RFPs still ignore them

Tomasz Tunguz (Theory Ventures)2h ago

Anthropic secretly tracked Claude users in China, then removed code after exposure

Ars Technica AI2h ago

Stay ahead with AI news

Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.

Get Started Free

Free · takes 30 seconds · unsubscribe anytime

1 minute a day. The AI essentials.

200+ sources · Email / LINE / Slack

Get it free →