Nex-N2.5-Pro: Agent Model That Self-Verifies Using Visual Feedback

This is an agent model that directly controls computers and browsers to perform long-duration tasks. It utilizes visual information not merely as input, but as an interface for environment perception and result verification. The core capability is its ability to view the screen, self-correct errors, and complete tasks.

It recorded 82.7 points on Terminal-Bench 2.1 and 82.2 points on OSWorld-Verified, demonstrating its coding and GUI manipulation capabilities. Inference depth can be adjusted via Thinking Modes settings, and it supports Function Calling for integration with external tools. It is compatible with major inference engines such as vLLM and SGLang.

Distributed under the Apache 2.0 license, it is available for commercial use. It is suitable for environments requiring sustained and precise agent tasks, such as scientific research or complex knowledge work. The Pro model runs on a single node with 8 H100 GPUs.