MiniMax M2.5

Production-scale agentic coding with full-stack development and office deliverables

Deploy model read docs

About model

MiniMax M2.5 is SOTA in coding, agentic tool use, search, and office work, extensively trained with reinforcement learning across 200,000+ complex real-world environments. The model achieves 80.2% SWE-Bench Verified while completing tasks 37% faster than M2.1, matching Claude Opus 4.6's speed. M2.5 exhibits architect-level planning capability, actively decomposing and planning features, structure, and UI design before writing code—spanning the entire development lifecycle from 0-to-1 system design through 90-to-100 comprehensive testing. Trained on 10+ programming languages across full-stack platforms (Web, Android, iOS, Windows), M2.5 delivers truly deliverable outputs in office scenarios on Together AI's production infrastructure.

SWE-Bench Verified

80.2%

SOTA coding across 200K+ real-world environments

Faster Than M2.1

37%

Matching Opus 4.6 speed with efficient decomposition

Real-World Training Environments

200K+

Model key capabilities

• Architect-Level Planning: Spec-writing with feature decomposition and UI design before coding—spanning 0-to-1 system design through 90-to-100 comprehensive testing
• SOTA Agentic Coding: 80.2% SWE-Bench Verified across 10+ languages and full-stack platforms—37% faster than M2.1, matching Opus 4.6 speed
• Office Deliverables: Word documents, PowerPoint presentations, Excel models trained with industry experts—59.0% win rate vs mainstream models
• Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

Quickstart guides

• RAG\ \ Building a RAG Workflow
• Agents\ \ Agent Workflows
• Apps\ \ Next.js Chat Quickstart

Model card

Architecture Overview:

• SOTA agentic model trained with reinforcement learning across 200,000+ complex real-world environments
• Forge agent-native RL framework with 40x training speedup through asynchronous scheduling and tree-structured sample merging
• CISPO algorithm ensuring MoE model stability during large-scale RL training
• Process reward mechanism for end-to-end generation quality monitoring in long-context agent rollouts
• Optimal trade-off between intelligence and response speed through trajectory-based task completion time evaluation
• Trained on 10+ programming languages: Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Ruby

Training Methodology:

• Extensive RL training in hundreds of thousands of real-world coding, search, and office work environments
• Collaboration with senior professionals in finance, law, and social sciences for office deliverables training
• Industry expert-designed requirements, feedback, and standards contributing to data construction
• Architect-level planning emerged during training: spec-writing before coding with feature decomposition
• Trained for efficient reasoning and optimal task decomposition reducing token consumption by 5% vs M2.1
• Full development lifecycle training: 0-to-1 system design, 1-to-10 development, 10-to-90 iteration, 90-to-100 testing

Performance Characteristics:

• Coding Excellence: 80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench, 79.7% Droid, 76.1% OpenCode
• Agentic Leadership: 76.3% BrowseComp (with context management), 20% fewer search rounds vs M2.1
• Office Deliverables: 59.0% win rate in GDPval-MM evaluation vs mainstream models
• Speed: 37% faster than M2.1 on SWE-Bench Verified (22.8 min vs 31.3 min), matching Claude Opus 4.6
• Cost Efficiency: 10% cost of Claude Opus 4.6 per task, $1/hour continuous operation at 100 TPS
• Token Efficiency: 3.52M tokens/task vs M2.1's 3.72M, 5% reduction through better decomposition
• Additional Benchmarks: 86.3% AIME25, 85.2% GPQA-D, 70.0% IFBench, 44.4% SciCode

Applications & use cases

Full-Stack Software Development:

• Architect-level planning: Spec-writing with feature decomposition, structure design, and UI planning before coding
• Complete development lifecycle: 0-to-1 system design and environment setup through 90-to-100 comprehensive testing
• 80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench across 10+ programming languages
• Full-stack platforms: Web, Android, iOS, Windows with server-side APIs, business logic, databases
• Complex system development beyond bug-fixing: feature iteration, code review, system testing
• Multi-environment generalization: 79.7% on Droid, 76.1% on OpenCode with different scaffoldings

Agentic Search & Tool Use:

• Industry-leading performance: 76.3% BrowseComp with context management
• Efficient decision-making: 20% fewer search rounds than M2.1 with better token efficiency
• Precise search rounds with optimal reasoning paths to results
• Stable performance across unfamiliar scaffolding environments
• Deep webpage exploration for information-dense professional tasks

Office Deliverables & Productivity:

• Word documents, PowerPoint presentations, Excel financial models as truly deliverable outputs
• Trained with senior professionals in finance, law, and social sciences
• 59.0% win rate vs mainstream models in GDPval-MM office work evaluation
• Industry-specific tacit knowledge integrated into training pipeline
• High-value workspace scenarios: financial modeling, legal documents, research reports
• Professional trajectory evaluation alongside deliverable quality assessment

Enterprise Coding Agents:

• Autonomous software development at production scale
• Multi-language, multi-platform development workflows
• Integration with Claude Code and major coding agent frameworks
• Repository-scale navigation, refactoring, and comprehensive testing
• Real-world deployment: 80% of MiniMax's newly committed code is M2.5-generated

Knowledge Work Automation:

• Automated research report generation with proper formatting
• Financial model creation following organizational standards
• Legal document preparation with industry compliance
• Presentation creation with professional design standards
• Real-world productivity: 30% of MiniMax company tasks autonomously completed by M2.5

Related models

MiniMax AI

• Model provider: MiniMax AI
• Main use cases: Chat, Function Calling
• Features: Function Calling, JSON Mode
• Speed: Medium
• Intelligence: Very High
• Deployment: Dedicated
• Parameters: 228.7B
• Context length: 192K
• Input price: $0.30 / 1M tokens, $0.06 (cached)/1M
• Output price: $1.20 / 1M tokens
• Input modalities: Text
• Output modalities: Text
• Released: February 11, 2026
• Last updated: February 14, 2026
• Quantization level: FP4
• External link: HuggingFace