MiniMaxAI/MiniMax-M2.5

MiniMax-M2.5 is a frontier-scale model positioned for software engineering and autonomous agent tasks, built on a trajectory of rapid iteration that included M2 and M2.1 predecessors. According to the official model card, it was trained with reinforcement learning across hundreds of thousands of complex real-world environments spanning more than ten programming languages including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby. This breadth of training environments underpins its "spec-writing tendency," a behavior where the model decomposes and plans features, structure, and UI design before producing code, mirroring how an experienced software architect would approach a task.

The model delivers competitive benchmark performance across coding, agentic tool use, search, and office productivity workloads. Independent reporting and the model card both cite 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp with context management, with the model reportedly outperforming Claude Opus 4.6 on the Droid harness (79.7% versus 78.9%) and matching its speed on SWE-Bench Verified while completing the evaluation 37% faster than M2.1. It is distributed under a Modified-MIT license and is hosted across multiple inference providers, with launch partners documenting its availability and offering comparison and benchmarking resources for application builders seeking cost-effective agentic coding capability.