Nex-N1: Agentic Models via Large-Scale Environment Construction

The founding work of the Nex-AGI ecosystem: a method to systematically scale the diversity and complexity of interactive environments for training autonomous agents — shifting agent learning from static imitation to incentive-driven decision making. Scaling is addressed along three orthogonal axes: Complexity via NexAU (a flexible framework for building agent hierarchies from simple configs), Diversity via NexA4A (auto-generating diverse agent hierarchies from natural language to cover open-ended domains), and Fidelity via NexGAP (a high-quality agentic data pipeline).

The resulting Nex-N1 agentic models are delivered as post-trains on open bases (DeepSeek-V3.1, Qwen3, internlm3), establishing the environment-construction + agentic-RL recipe later carried into Nex-N2.