Nex-N2-Pro is an open-weight agentic mixture-of-experts (MoE) language model developed by Nex AGI. Released in June 2026, it is built on the Qwen 3.5 architecture and contains a total of 397 billion parameters, with 17 billion parameters active per token during inference. The model is specifically engineered for real-world productivity scenarios, emphasizing "Agentic Thinking" to handle complex, long-horizon tasks that require planning, environment feedback, and iterative debugging.
The model's architecture incorporates two specialized reasoning layers: Adaptive Thinking and Coherent Thinking. Adaptive Thinking enables the model to dynamically regulate its reasoning depth based on the complexity of the prompt, allowing for rapid execution of simple tasks and deep deliberation for critical decisions. Coherent Thinking maintains a consistent reasoning paradigm across different modalities and agentic tasks, ensuring stable capability transfer between general reasoning and specialized domains like coding or research.
Nex-N2-Pro is multimodal, accepting both text and image inputs, and supports a context window of 262,144 tokens. It has demonstrated high performance on agentic benchmarks, notably scoring 80.8 on SWE-Bench Verified and 75.3 on Terminal-Bench 2.1. These results suggest a high degree of proficiency in autonomous software engineering, tool use, and terminal-based execution compared to other open-weight models.
For effective integration into agentic workflows, the developers recommend using specific parsers, such as the qwen3_coder flag for tool calling and the qwen3 flag for reasoning traces. Preferred sampling parameters include a temperature of 0.7, a top_p of 0.95, and a top_k of 40. The model is distributed under the Apache 2.0 license, making it available for a wide range of applications.