MiniMax M2.5

• 🆕 MiniMax's productivity-focused LLM optimized for coding and agentic workflows.
• 🧠 Trained via large-scale RL across 200,000+ real-world environments.
• 📏 Catalog lists a 198K-token context window for long tasks.
• 🔧 Supports function calling, web search, and multi-step tool use.
• 🎯 Vendor-reported 80.2% on SWE-Bench Verified, 51.3% Multi-SWE-Bench.
• ⚡ Completes SWE-Bench Verified roughly 37% faster than M2.1.
• 📚 "Spec-writing" behavior: plans architecture before writing code.
• 💬 Strong on office workflows like Word, PowerPoint, Excel modeling.

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax M2.5, released in February 2026, is a text model from MiniMax built for coding, agentic tool use, and office productivity. According to MiniMax, it was extensively trained with reinforcement learning across more than 200,000 complex real-world environments using the company's Forge agent-native RL framework and CISPO algorithm, with a process-reward mechanism for monitoring generation quality in long-context agent rollouts. The catalog lists a 198K-token context window plus reasoning, code-optimization, function-calling, and web-search capabilities.

Against its same-family predecessor M2.1, MiniMax reports concrete gains. On the provider's reported SWE-Bench Verified, M2.5 scores 80.2% while completing the evaluation about 37% faster than M2.1—end-to-end runtime dropping from 31.3 to 22.8 minutes and tokens per task falling from 3.72M to 3.52M. MiniMax also reports 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp with context management. A notable behavioral change is M2.5's tendency to decompose and plan features, structure, and UI like a software architect before coding.

MiniMax positions M2.5 for the full development lifecycle across Web, Android, iOS, Windows, and Mac, and for workspace tasks such as financial modeling and report generation. A higher-throughput M2.5-highspeed variant is also offered.

M2.5 was later succeeded within the family by MiniMax M2.7, MiniMax M3, and MiniMax M3 Preview, all sharing the same coding-and-agentic focus. For deployment, MiniMax recommends vLLM or SGLang.