MiMo-V2.6-Pro-RL: TerminalBench 89.9, Agent Coding Specialized Model

Achieved a score of 89.9 on the TerminalBench 2.1 benchmark, demonstrating the ability to control complex terminal environments. It also scored 71.9 on the Deep Swe benchmark, showing high performance in real-world software development scenarios. With a Toolathlon Verified evaluation score of 76.9, external tool invocation and agent task execution are implemented reliably.

It features a multimodal architecture capable of processing text, images, audio, and video inputs. It supports long contexts, allowing for the analysis of vast codebases or entire documents at once. With built-in thinking mode and tool call capabilities, it is suitable for agent applications requiring complex logical reasoning and external API integration.

With approximately 524B parameters, it supports FP8 and 8-bit quantization formats, ensuring deployment flexibility. Under the MIT license, it can be adopted for commercial projects without restrictions. Optimized for English and Chinese processing, it is well-suited for developing coding agents or multimodal chatbots.