MiniMax M2 on Novita AI: Compact Frontier Intelligence for Agentic Coding and Multi-Step Reasoning

MiniMax M2 is now available on the Novita AI platform, bringing compact frontier intelligence to agentic coding workflows and multi-step task execution. With 10 billion activated parameters out of 230 billion total, this high-efficiency model delivers near-frontier performance while maintaining exceptional speed and cost efficiency that makes advanced AI accessible to more developers and teams.

This latest addition to Novita AI’s model collection excels in end-to-end coding tasks, multi-file editing, compile-run-fix loops, and test-validated repair. Whether you’re building developer assistants, scaling agent-driven applications, or creating reasoning-intensive tools, MiniMax M2 delivers the intelligence you need with the responsiveness you demand through Novita AI’s developer-friendly infrastructure.

What is MiniMax M2?

MiniMax M2 is a compact, high-efficiency large language model optimized specifically for end-to-end coding and agentic workflows. Built with a Mixture of Experts (MoE) architecture featuring 10 billion activated parameters out of 230 billion total, it achieves near-frontier intelligence across general reasoning, tool use, and multi-step task execution while maintaining remarkably low latency and deployment efficiency.

Optimized for Agentic Coding

MiniMax M2 demonstrates exceptional performance in code generation, multi-file editing, compile-run-fix loops, and test-validated repair. The model is designed from the ground up for autonomous coding agents that need to understand complex codebases, make intelligent decisions across multiple files, and recover gracefully from execution errors.

Superior Efficiency Architecture

The model’s small activation footprint enables fast inference, high concurrency, and improved unit economics. With only 10B active parameters processing each request, MiniMax M2 delivers frontier-level intelligence at a fraction of the computational cost, making it ideal for large-scale deployments where responsiveness and cost efficiency matter.

Comprehensive Intelligence Range

Beyond coding, MiniMax M2 performs competitively in agentic evaluations spanning mathematics, science, instruction-following, and multi-step reasoning. The model handles long-horizon planning, complex retrieval tasks, and error recovery effectively, making it suitable for reasoning-driven applications across diverse domains.

Key Features and Capabilities

Code Generation and Software Engineering

MiniMax M2 shows strong results on rigorous coding benchmarks including SWE-Bench Verified, Multi-SWE-Bench, and Terminal-Bench. The model excels at:

• Multi-file editing: Navigate and modify complex codebases spanning multiple files with contextual awareness
• Compile-run-fix loops: Iteratively debug and repair code based on compiler feedback and runtime errors
• Test-validated repair: Generate fixes that pass existing test suites while maintaining code quality
• End-to-end task completion: Handle complete development workflows from specification to working implementation

Agentic Capabilities

MiniMax M2 performs effectively in agentic evaluations such as BrowseComp and GAIA, demonstrating proficiency in:

• Long-horizon planning: Break down complex objectives into actionable steps and execute multi-stage plans
• Tool use and integration: Leverage external tools, APIs, and systems within reasoning workflows
• Execution error recovery: Detect failures, diagnose issues, and adapt strategies to reach objectives
• Multi-step reasoning: Maintain coherent reasoning chains across extended interactions

Composite Intelligence

Benchmarked by Artificial Analysis, MiniMax M2 ranks among the top open-source models for composite intelligence. This comprehensive evaluation spans:

• Mathematics: Solve complex mathematical problems requiring multi-step reasoning
• Science: Apply scientific knowledge and reasoning to domain-specific challenges
• Instruction-following: Accurately interpret and execute detailed, multi-faceted instructions
• General reasoning: Demonstrate flexible problem-solving across diverse scenarios

The model’s balanced performance across these dimensions makes it suitable for applications requiring versatile intelligence rather than narrow specialization.

Performance Benchmarks

Artificial Analysis Composite Intelligence Ranking

According to Artificial Analysis benchmarking, MiniMax M2 ranks among the top open-source models for composite intelligence, spanning mathematics, science, and instruction-following. This comprehensive evaluation provides a holistic view of the model’s capabilities across multiple dimensions.

Software Engineering Benchmarks

MiniMax M2 demonstrates strong results on rigorous software engineering evaluations:

SWE-Bench Verified: The model shows strong performance on this challenging benchmark that evaluates the ability to resolve real-world GitHub issues from popular open-source repositories.

Multi-SWE-Bench: MiniMax M2 demonstrates effective performance on this benchmark that assesses handling of multiple software engineering tasks.

Terminal-Bench: The model performs well on this evaluation measuring command-line interaction and system-level task completion capabilities.

These results demonstrate MiniMax M2’s readiness for practical software engineering applications including code generation, multi-file editing, compile-run-fix loops, and test-validated repair.

Agentic Evaluation Performance

MiniMax M2 performs competitively in agentic evaluations such as BrowseComp and GAIA. These benchmarks assess the model’s ability to effectively handle:

• Long-horizon planning
• Information retrieval across multiple sources
• Recovery from execution errors
• Multi-step reasoning and task completion

The model’s performance in these evaluations demonstrates its suitability for autonomous agent applications and reasoning-driven workflows.

Getting Started with MiniMax M2 on Novita AI Platform

Novita AI offers multiple ways to access MiniMax M2, designed for different skill levels and use cases.

Use the Playground (No Coding Required)

Sign up and start experimenting with MiniMax M2 in seconds through an interactive interface. Test prompts, see outputs in real-time, and compare MiniMax M2 with other leading models. Perfect for prototyping and understanding what the model can do before building full implementations.

Integrate via API (For Developers)

Connect MiniMax M2 to your applications using Novita AI’s unified REST API.

Direct API Integration (Python Example)

from openai import OpenAI
  
client = OpenAI(
    base_url="https://api.novita.ai/openai",
    api_key="",
)

model = "minimax/minimax-m2"
stream = True # or False
max_tokens = 65536
system_content = "Be a helpful assistant"
temperature = 1
top_p = 1
min_p = 0
top_k = 50
presence_penalty = 0
frequency_penalty = 0
repetition_penalty = 1
response_format = { "type": "text" }

chat_completion_res = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "system",
            "content": system_content,
        },
        {
            "role": "user",
            "content": "Hi there!",
        }
    ],
    stream=stream,
    max_tokens=max_tokens,
    temperature=temperature,
    top_p=top_p,
    presence_penalty=presence_penalty,
    frequency_penalty=frequency_penalty,
    response_format=response_format,
    extra_body={
      "top_k": top_k,
      "repetition_penalty": repetition_penalty,
      "min_p": min_p
    }
  )

if stream:
    for chunk in chat_completion_res:
        print(chunk.choices[0].delta.content or "", end="")
else:
    print(chat_completion_res.choices[0].message.content)

Multi-Agent Workflows with OpenAI Agents SDK

Build sophisticated multi-agent systems with plug-and-play integration, support for handoffs, routing, and tool integration.

Connect with Third-Party Platforms

Coding Agents: Integrate with popular coding assistants like Claude Code, Cursor, Windsurf, Cline, and Aider through OpenAI-compatible APIs and Anthropic-compatible APIs.

Orchestration Frameworks: Connect with LangChain, Dify, CrewAI, and Langflow using official connectors.

Hugging Face: Novita AI is an official inference provider for Hugging Face, ensuring broad ecosystem compatibility.

Conclusion

MiniMax M2 on Novita AI delivers compact frontier intelligence optimized for agentic coding and multi-step reasoning tasks. With 10 billion activated parameters (from 230 billion total) achieving near-frontier performance and exceptional efficiency metrics, MiniMax M2 represents a breakthrough in accessible advanced AI capabilities.

The model’s strong performance on SWE-Bench Verified, Multi-SWE-Bench, Terminal-Bench, BrowseComp, and GAIA demonstrates its readiness for production coding agents, developer assistants, and reasoning-driven applications. Benchmarked by Artificial Analysis among the top open-source models for composite intelligence spanning mathematics, science, and instruction-following, MiniMax M2 offers the intelligence and efficiency you need for large-scale deployments.

Whether you’re building autonomous software engineering tools, scaling multi-agent systems, or creating intelligent automation, MiniMax M2’s small activation footprint enables fast inference, high concurrency, and improved unit economics.

Start exploring MiniMax M2 today through Novita AI’s Playground, integrate it via API, or connect through third-party platforms to enhance your development workflow with powerful coding assistance, multi-step reasoning, and cost-effective frontier intelligence.

Novita AI is an AI cloud platform that offers developers an easy way to deploy AI models using our simple API, while also providing affordable and reliable GPU cloud for building and scaling.