MiniMax-M2.5: A Mixture-of-Experts Model
TL;DR
MiniMax-M2.5 is a 229B-parameter Mixture-of-Experts model built for coding, tool use, professional search, and office productivity. It combines a 196,608-token context window with reinforcement learning across hundreds of thousands of real-world environments.
What the Model Is
MiniMax-M2.5, now listed on AIOZ AI, is the latest model in MiniMax’s M2 series.
The model is optimized for practical agentic tasks that require planning, tool interaction, and complete deliverables. Its target workflows range from full-stack development and technical research to PowerPoint creation and Excel financial modeling.
How the Model Architecture Works
Its routing system activates 8 of 256 experts, providing substantial model capacity with less computation than a similarly sized dense model.
The architecture includes:
- 62 transformer layers
- 48 attention heads
- 8 key-value heads
- 196,608-token context support
- Multi-Token Prediction
- FP8 weight optimization
Grouped Query Attention reduces KV-cache requirements during long interactions, while the extended context supports large codebases, documents, research histories, and multi-step tool trajectories.
How the Model Is Trained
MiniMax-M2.5 was trained using Forge, MiniMax’s agent-native reinforcement learning framework.
Forge separates agent logic from the underlying training and inference engines, allowing different tools, memory systems, and agent scaffolds to participate in the same training infrastructure.
The framework uses asynchronous scheduling and structured sample merging to improve training throughput.
Its core reinforcement learning algorithm, CISPO, is designed to maintain training stability for large Mixture-of-Experts models and long agent trajectories.
Core Capabilities
- Software development: Supports end-to-end development across 200,000+ environments and 10+ programming languages.
- Search and tool calling: Conducts multi-step research and tool-based workflows using approximately 20% fewer search rounds than M2.1.
- Office productivity: Produces Word, PowerPoint, and Excel deliverables with a reported 59% win rate on GDPval-MM.
Performance and Efficiency
- 80.2% on SWE-Bench Verified
- 51.3% on Multi-SWE-Bench
- 76.3% on BrowseComp with context management
- 86.3 on AIME25
- 85.2 on GPQA-D
- 70.0 on IFBench
On SWE-Bench Verified, the average runtime decreased from 31.3 minutes with M2.1 to 22.8 minutes with M2.5, representing a 37% improvement.
MiniMax attributes this gain to stronger task decomposition and improved token efficiency. Its high-speed serving configuration reaches up to 100 output tokens per second.
Key Technical Details
- Model ID: MiniMax-M2.5
- Model type: Mixture-of-Experts language model
- Total parameters: 229B
- Transformer layers: 62
- Available experts: 256
- Experts selected per token: 8
- Context window: 196,608 tokens
- Attention: Grouped Query Attention
- Attention heads: 48
- Key-value heads: 8
- Training framework: Forge
- Reinforcement learning algorithm: CISPO
- Quantization configuration: FP8
- License: MiniMax Model License
Download It on AIOZ AI
229 billion parameters, a 196,608-token context window, and agent-native reinforcement learning: MiniMax-M2.5 is built to move from instructions to complete deliverables.
Download MiniMax-M2.5 on AIOZ AI and start building coding, research, and professional productivity agents.
FAQ
Q1: What is MiniMax-M2.5?
It is a 229B-parameter Mixture-of-Experts model.
Q2: How much context does MiniMax-M2.5 support?
Its published configuration supports up to 196,608 tokens.
Q3: What does its Mixture-of-Experts architecture do?
The model routes each token through 8 of its 256 experts, reducing active computation while retaining access to the capacity of a much larger model.