What is nex-n2?

Nex-N2 is a pair of open-source AI models designed to handle complex, multi-step tasks on a computer, like writing and running code, browsing the web, using external tools, and fixing errors in a continuous loop. The two variants are Nex-N2-Pro and Nex-N2-mini: the Pro version is larger and more capable, while the mini version is smaller and faster, both built on top of Qwen3.5, an existing open-source model family. The project comes from a company called Nex-AGI, and the central idea is called Agentic Thinking. Rather than treating reasoning and action as separate steps, the model keeps them connected in a single loop: it understands what you want, plans steps, writes code, runs it, reads the results, fixes mistakes, and repeats until the job is done. It includes two design choices to make this work: one lets the model decide when it needs to think carefully versus when it can act quickly, and another keeps its reasoning style consistent across different types of tasks. According to the benchmark comparisons in the README, Nex-N2-Pro scores competitively against well-known AI models on coding challenges and on tasks that require browsing and tool use. The mini version is smaller but still performs well on many of the same tasks. Running either model requires significant hardware: the Pro variant is designed for a cluster of 16 high-end GPUs spread across two servers, while the mini needs at least two high-end GPUs on one server. Installation involves a custom fork of a serving framework called sglang, or a prebuilt Docker image the team provides. Once running, it exposes an OpenAI-compatible API endpoint, so developers can connect to it using standard tools. The models are available to download from Hugging Face and ModelScope.