Available Models

MiniMax-M2 offers production-ready models optimized for different development workflows. This page documents stable profiles (M2, M2.1, M2.5) plus deployment-configurable support for M2.7.

For implementation-accurate M2.7 behavior (aliases, fallback, env-based pricing), see the MiniMax-M2.7 Guide.

Quick Comparison

MiniMax-M2

Model ID: MiniMax-M2 Released: October 27, 2025

Strengths

• ✅ Agent-First Design: Reliable coordination of Shell, Browser, Python interpreter, and MCP tools
• ✅ Tool Integration: Industry-leading tool-use performance, matches top overseas models
• ✅ Deep Search: Near best-in-class research and information retrieval capabilities
• ✅ Production Stability: Internally validated by MiniMax engineering teams before release
• ✅ Fast Inference: ~100 tokens/second throughput for responsive interactions
• ✅ Cost Efficiency: Balanced intelligence, speed, and affordability

Ideal Use Cases

• Python and JavaScript-heavy agent workflows
• Cost-sensitive production deployments
• Shell automation and CLI tool orchestration
• Browser automation and web scraping tasks
• General-purpose coding assistance
• Teams prioritizing inference speed over specialized language support

Supported Platforms

• OpenAI-compatible Chat Completions API
• Claude-compatible Messages API
• Claude Code, Cline, Kilo Code, RooCode, BlackBox
• Factory AI's Droid

MiniMax-M2.1

Model ID: MiniMax-M2.1 Released: December 23, 2025

Major Improvements Over M2

🌍 Systematically Enhanced Multi-Language Programming

Strong performance across Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, JavaScript, and more. Achieves industry-leading multilingual capabilities, surpassing Claude Sonnet 4.5 and approaching Claude Opus 4.5 in multi-language scenarios.

⚡ Efficiency Revolution

Generates more concise responses and thought chains, significantly reducing token consumption while improving response speed. Delivers higher quality with lower operational costs.

📱 Native Mobile Development

Substantially strengthened native Android and iOS development capabilities. Enhanced understanding of platform-specific patterns, frameworks, and best practices.

🎨 Design & Aesthetic Comprehension

Improved aesthetic judgment for web applications, better alignment with modern design principles and user experience standards.

🏢 Office & Composite Workflows

Enhanced handling of composite instruction constraints for office scenarios, supporting complex multi-step business processes.

Benchmark Highlights

• VIBE Aggregate: 88.6 (Visual & Interactive Benchmark for Evaluation)
  • VIBE-Web: 91.5
  • VIBE-Android: 89.7
• SWE-bench Verified: Strong framework generalization across diverse agent tooling
• Multilingual Scenarios: Outperforms Claude Sonnet 4.5, approaches Claude Opus 4.5

Ideal Use Cases

• Multi-language codebases (Rust, Java, Golang, C++, etc.)
• Native Android and iOS application development
• Web applications requiring strong design sensibility
• Complex office automation and business workflows
• Teams prioritizing output quality and token efficiency
• Projects requiring broad framework and language support

Supported Platforms

• OpenAI-compatible Chat Completions API
• Claude-compatible Messages API
• Claude Code, Cline, Kilo Code, RooCode, BlackBox
• Factory AI's Droid
• GUI automation via text commands

MiniMax-M2.5

Model ID: MiniMax-M2.5 Released: February 12, 2026

Major Improvements Over M2.1

🏗️ Architect-Style Planning

M2.5 actively decomposes and plans features, structure, and UI design before writing any code. This spec-writing tendency emerged during training — the model thinks like an experienced software architect.

💻 SOTA Coding Performance

Trained on 13+ languages (Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Ruby) across 200,000+ real-world environments. Covers the entire development lifecycle: 0-to-1 system design, 1-to-10 development, 10-to-90 feature iteration, and 90-to-100 code review and testing. Full-stack across Web, Android, iOS, and Windows.

• SWE-Bench Verified: 80.2% (SOTA)
• Multi-SWE-Bench: 51.3%
• On Droid: 79.7 (> Opus 4.6's 78.9)
• On OpenCode: 76.1 (> Opus 4.6's 75.9)
• 37% faster than M2.1 on SWE-Bench, matching Claude Opus 4.6 speed

🔍 Industry-Leading Search & Tool Calling

Best-in-class performance on BrowseComp (76.3% with context management) and Wide Search. Uses ~20% fewer rounds than M2.1 to achieve better results — more efficient reasoning paths, not just correct answers.

📊 Office & Finance Workflows

Excels at Excel-based tasks (MEWC benchmark), financial modeling, and complex multi-step business processes. Handles composite instruction constraints for real enterprise scenarios.

⚡ Speed & Cost

Completes SWE-Bench Verified 37% faster than M2.1. It costs just $1 to run continuously for an hour at 100 tokens/s, or $0.30 at 50 tokens/s.

Benchmark Highlights

• SWE-Bench Verified: 80.2% (SOTA)
• Multi-SWE-Bench: 51.3%
• BrowseComp: 76.3% (with context management)
• VIBE-Pro: On par with Opus 4.5 (upgraded from VIBE to more complex Pro version)
• Terminal Bench 2: Industry-leading
• Context Window: 200K tokens

Ideal Use Cases

• Complex full-stack development across Web, Android, iOS, and Windows
• Large-scale codebase modifications and system-level refactoring
• Agentic workflows requiring deep search and multi-step tool use
• Enterprise office automation and financial modeling
• Projects requiring architect-level planning before implementation
• Long-context tasks benefiting from the 200K token window

Supported Platforms

• OpenAI-compatible Chat Completions API
• Claude-compatible Messages API
• Claude Code, Cline, Kilo Code, RooCode, BlackBox
• Factory AI's Droid
• MiniMax Agent (10,000+ Experts built on M2.5)

MiniMax-M2.7

Model ID: MiniMax-M2.7

For endpoint-level behavior and rollout details, read MiniMax-M2.7 Guide.

What is fixed

• Supported on both POST /api/v1/chat/completions and POST /api/v1/messages
• Uses the same request/response schema as other MiniMax models
• Has its own model identifier for usage records and billing attribution
• Supports alias-based routing (codex-MiniMax-M2.7, MiniMax M2.7)

What is deployment-configurable

• Effective pricing
• Context window and capability profile
• Availability rollout state
• API strictness behavior (MINIMAX_MODEL_STRICT_API)

To inspect your live environment values, call GET /api/v1/models.

How to Choose

Choose MiniMax-M2 if you:

• Primarily work with Python and JavaScript
• Need fast, stable agent workflows with proven reliability
• Prioritize inference speed for interactive applications
• Run cost-sensitive production workloads

Choose MiniMax-M2.1 if you:

• Develop in Rust, Java, Golang, C++, Kotlin, or Objective-C
• Build native Android or iOS applications
• Require higher output quality with lower token costs
• Work on web applications where design matters
• Handle complex multi-step business processes

Choose MiniMax-M2.5 if you:

• Need SOTA coding performance across 13+ languages
• Work on complex full-stack projects spanning multiple platforms
• Require architect-level planning and spec-writing before implementation
• Run agentic workflows with deep search and multi-step tool use
• Handle enterprise office automation or financial modeling
• Need the largest context window (200K tokens)

Choose MiniMax-M2.7 if you:

• Need the newest deployment-specific MiniMax profile without changing integration code
• Want independent billing attribution for M2.7 traffic
• Plan to tune pricing or rollout strategy separately from M2/M2.1/M2.5

How to Specify Models

Pass the model field in your API request:

Chat Completions (OpenAI format)

{
  "model": "MiniMax-M2.5",
  "messages": [
    { "role": "user", "content": "Write a Rust function for async HTTP requests" }
  ]
}

Messages API (Claude format)

{
  "model": "MiniMax-M2.5",
  "system": "You are an expert Android developer.",
  "messages": [
    { "role": "user", "content": "Implement a Material 3 bottom sheet" }
  ],
  "max_tokens": 2048
}

Default Behavior

Defaults are channel-specific:

• Web chat UI defaults to MiniMax-M2.5
• OpenAI-compatible API (POST /api/v1/chat/completions) defaults to MiniMax-M2 when model is omitted
• Messages API (POST /api/v1/messages) defaults to MiniMax-M2 when model is omitted

This preserves backward compatibility for existing API clients while keeping the hosted chat experience on M2.5.

Pricing

Baseline launch-window pricing for MiniMax-M2, MiniMax-M2.1, and MiniMax-M2.5:

MiniMax-M2.7 may use a different price profile depending on deployment configuration. Always check GET /api/v1/models for effective model-level rates before cost planning.

Usage charges appear in your billing dashboard with per-model breakdowns, allowing you to track consumption across M2, M2.1, M2.5, and M2.7 independently.

Recommended Parameters

For optimal performance with all models:

{
  "temperature": 1.0,
  "top_p": 0.95,
  "top_k": 40
}

Adjust temperature downward (0.3–0.7) for deterministic outputs in compliance-sensitive or production-critical scenarios.

Access Methods

API Integration:

• OpenAI SDK: Set base_url to https://minimax-m2.com/api/v1/
• Anthropic SDK: Set base_url to https://minimax-m2.com/api/

Direct API:

• Chat Completions Endpoint
• Messages Endpoint

Developer Tools:

• Claude Code Configuration
• Codex Integration

Migration Notes

Switching to M2.7

1. Update the model field: Change your request model to "MiniMax-M2.7"
2. No endpoint changes required: OpenAI-compatible and Claude-compatible contracts remain unchanged
3. Confirm live pricing/capabilities: Query GET /api/v1/models in your deployment before rollout
4. Roll out safely: Start with a subset of traffic, compare quality/cost, then expand

Switching to M2.5

1. Update the model field: Change "MiniMax-M2.1" to "MiniMax-M2.5" in your request body
2. No API changes required: All models support identical endpoint contracts
3. Leverage 200K context: M2.5 supports 200K tokens, up from 128K in M2/M2.1
4. Expect architect-style output: M2.5 may produce spec/plan documents before writing code — this is by design
5. Monitor speed improvements: M2.5 completes agentic tasks 37% faster than M2.1

Switching from M2 to M2.1

1. Update the model field: Change "MiniMax-M2" to "MiniMax-M2.1" in your request body
2. No API changes required: Both models support identical endpoint contracts
3. Monitor token usage: M2.1's concise responses may reduce your token consumption
4. Test prompts: M2.1's enhanced capabilities may produce different (often better) outputs for the same inputs

Backward Compatibility

All M2 prompts, system messages, and tool configurations work unchanged with M2.1, M2.5, and M2.7. You can switch models without modifying application logic.

Official Resources

• M2.7 Model Page: https://www.minimax.io/models/text/m27
• M2.7 Report: https://www.minimax.io/news/minimax-m27-en
• M2.5 Announcement: https://www.minimax.io/news/minimax-m25
• M2.1 Announcement: https://www.minimax.io/news/minimax-m21
• M2 Announcement: https://www.minimax.io/news/minimax-m2
• API Documentation: https://platform.minimax.io/docs/guides/text-generation
• Agent Product: https://agent.minimax.io/
• Open-source Weights: HuggingFace (MiniMaxAI organization)