M2.1: Multilingual and Multi-Task Coding with Strong Generalization (www.minimaxi.com)

MiniMax-M2.1 has made a substantial advancement in coding capabilities, positioning itself alongside top-tier models in various benchmarks. As an open-source AI model tailored for agentic scenarios, it excels in code generation, tool utilization, and long-term planning. The introduction of MultiMax-M2.1 comes in response to the limitations of SWE-Bench, which primarily focuses on bug-fixing in Python. The new model addresses these challenges by incorporating a comprehensive training system that spans over ten programming languages and supports multi-task functionalities essential for real-world software development, including code reviews, performance optimizations, and test generation. One of the notable technical achievements of MiniMax-M2.1 is its capability to manage a diverse set of coding environments and adapt to varying scaffolds used by developers. To facilitate this, a high-concurrency sandbox infrastructure was developed, capable of executing thousands of environments simultaneously, thus supporting efficient training across multiple languages. Additionally, the model has shown exceptional performance in understanding complex instruction constraints and handling contextual management strategies, improving its adaptability and instruction-following capabilities. Through these enhancements, MiniMax-M2.1 has not only bridged the gap between artificial and human coding but has also laid the groundwork for ongoing advancements in coding agents.