A new multimodal AI agent can handle text, image, and video inputs to navigate software interfaces and complete multi-step digital tasks. The Nex-N2.6 model, which was released as open-source by the Shanghai Innovation Institute, achieved a 100% valid tool selection rate across 956 tested contexts, a result matching the performance of GPT-6-Sol.
In performance tests, Nex-N2.6 scored 84.1 on Terminal-Bench 2.1 and 83.0 on OSWorld, exceeding the 82.7 and 71.7 scores of DeepSeek-V4.1-Flash, respectively. The update also reduced the model’s reasoning repetition rate to 1.6% from 16.0% in the preceding Nex-N2.5-Pro version. The tool is currently available via HuggingFace, GitHub, and the OpenRouter and MOSS APIs.
