Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

Abstract: The evolution of LLMs from passive responders to autonomous agents necessitates a fundamental shift in learning paradigms -- from static imitation to incentive-driven decision making. However, this transition is significantly impeded by the lack of scalable infrastructure capable of constructing high-quality interaction signals for effective policy learning. To address this, we introduce a comprehensive method designed to systematically scale the diversity and complexity of interactive environments. Our method realizes this scaling by addressing three orthogonal dimensions: (1) Complexity: NexAU, a flexible agent framework that supports building complex agent hierarchies via simple configurations; (2) Diversity: NexA4A automatically generates diverse agent hierarchies from natural language to cover infinite domains; and (3) Fidelity: NexGAP bridges the simulation-reality gap by integrating dynamic real-world environment for grounded trajectories synthesis. We train Nex-N1 upon the diverse and complex interactive environments established by our infrastructure. Empirical results on benchmarks such as SWE-bench and tau2 demonstrate that Nex-N1 consistently outperforms SOTA open-source models and achieves competitive performance against frontier proprietary models on complex agentic tasks. We open-source the Nex ecosystem and model weights to facilitate further research.

• The paper introduces a modular ecosystem (NexAU, NexA4A, NexGAP) that automates agent and environment synthesis.
• The paper demonstrates state-of-the-art performance, surpassing GPT-5 on tool use, coding benchmarks, and cross-framework tasks.
• The paper presents a unified infrastructure that enables scalable generation of realistic, multimodal, and multilingual agent trajectories.

Nex-N1: A Unified Ecosystem for Agentic Model Training and Large-Scale Environment Construction

Introduction

The development of agentic LLMs requires rethinking both objective functions and data generation, transitioning from static next-token prediction to interactive, goal- and feedback-driven operation. Nex-N1 is a suite of models trained within a comprehensive, scalable ecosystem comprising automated agent construction, orchestration, and data generation infrastructure. This paper introduces three pillars—NexAU, NexA4A, and NexGAP—that collectively enable the generation of diverse, complex, and realistic agentic environments at scale, thereby driving robust agentic capabilities.

Unified Infrastructure for Agentic Scaling

NexAU: Modular Agent Runtime

NexAU defines a flexible, high-throughput agent runtime that abstracts heterogeneous agent frameworks into a unified, recursive architecture. It features hierarchical agent decomposition, dynamic tool/sub-agent composition, and context/state isolation, all programmable via declarative YAML schemas.

Figure 1: A simplified NexAU definition. Sub-agents are composed alongside standard tools, enabling arbitrary depth of composition.

This recursive structure allows the simulation of nested and complex interaction topologies, with parent and child agents operating isolated reasoning and ReAct loops. The abstraction of tools and sub-agents as fungible callable entities enables arbitrarily deep compositions, supporting both classical and novel agentic paradigms.

NexA4A: Generative Synthesis of Agents and Frameworks

NexA4A automates the generation of diverse agents and multi-layer frameworks from natural language specifications. It hierarchically decomposes task-level agent requirements into system prompts, meta-skills, tool integrations, and sub-agent instantiations. Its FrameworkBuilder and AgentBuilder components enable procedural synthesis of agent population graphs with highly variable node depth and interaction structure.

Figure 2: NexA4A agent framework workflow: a comprehensive architecture diagram from framework description to high-quality trajectory generation.

This generative approach completely decouples agent design from manual coding, systematically expanding the space of encountered agentic tasks and workflows for training, while maintaining strict compatibility with the NexAU runtime substrate.

NexGAP: Authentic and Heterogeneous Trajectory Generation

NexGAP coordinates end-to-end agent-environment execution, leveraging both synthetic and real-world Model Context Protocol (MCP) tools to generate grounded, multi-format agentic trajectories. The system curates over one hundred high-fidelity MCP tools (APIs, databases) and synthesizes multilingual, multi-difficulty task queries using an inverse-frequency weighted taxonomy.

It implements dynamic supervisor tools and multimodal feedback mechanisms for quality assurance, including iterative code repair and objective filtering of reward-hacking or hallucination behaviors. For non-agent data, search grounding and agentic pipelines further enhance data realism and reduce error rates.

Empirical Results and Evaluation

Benchmark Performance

Nex-N1 achieves state-of-the-art results among open-source models across a spectrum of agentic benchmarks, and is competitive with proprietary models in areas such as long-horizon planning, coding, and API-based tool use. Benchmarks include $\tau^2$ (constraint-solving), GAIA2 (agentic scenarios), SWE-bench (code repair), Terminal Bench 2, BaxBench (backend application development), and the Berkeley Function Calling Leaderboard (BFCL v4).

Figure 3: Performance comparison of Nex-N1 against state-of-the-art models on agent and coding benchmarks.

In tool usage and agentic coding, the largest Nex-N1 model surpasses GPT-5 on multiple open benchmarks. Qwen3 and DeepSeek models fine-tuned with the Nex infrastructure demonstrate step-function improvements over their non-Nex-N1 counterparts, indicating strong transfer and robustness of the generated data and agentic training regimen.

Human Evaluation: Agentic Coding and Web Creation

Human evaluators rate Nex-N1 as matching or outcompeting strong proprietary baselines (claude-sonnet-4.5, Minimax-M2) in realistic coding tasks involving end-to-end project workflows, functional correctness, and execution robustness.

Figure 4: Human evaluation results on agentic coding.

For web page synthesis, Nex-N1 consistently achieves high visual quality and completeness, outperforming all open-source and most proprietary competitors.

Figure 5: Human evaluation results on webpage creation.

Deep Research and Information Visualization

A dedicated research agent, built on Nex-N1, integrates web search, multi-step planning, and multimodal report generation. On the Deep Research Benchmark, the agent scores $47.0\%$, evidencing versatile performance across information acquisition and synthesis tasks.

Figure 6: Result of our deep research agent on the deep research benchmark.

The system supports both text-centric and visually-rich formats for downstream communication, supporting research pipelines which require high-fidelity, human-consumable outputs.

Figure 7: Demo of our deep research agent.

Poster generation from scientific papers is demonstrated using integrated PDF parsing, visual asset retrieval, bilingual rendering, and iterative feedback for layout/perceptual refinement.

Figure 8: Demo of our Paper2Poster agent (adapted from~\citep{xi2025bapo}).

Cross-Framework Robustness

Nex-N1's agentic capabilities generalize stably across diverse agent frameworks (OpenHands, Claude Code, Terminus 2), outperforming or matching other large-scale models in the SWE-bench subset. This confirms framework-agnostic reasoning and execution, attesting to the value of infrastructural abstraction and data normalization.

Implications and Future Directions

By automating the construction and diversification of interactive agents and their data, Nex-N1 challenges the field's dependence on manually-designed environments and static datasets. Its ecosystem enables systematic scaling along complexity, diversity, and reality-fidelity axes, facilitating research into compositional, long-horizon, and robustness properties of agentic models.

The strong numerical results, particularly the findings that Nex-N1 surpasses GPT-5 on tool use among open-source models and shows robust cross-framework generalization, indicate a path toward scalable, verifiable agent learning systems.

Future developments are directed toward extending this infrastructure as a large-scale, fully automated simulation platform for long-horizon reinforcement learning, incorporating dynamic environment generation, objective difficulty scaling, and self-improving agent populations. This lays the groundwork for training universal agents in increasingly realistic, open-ended, and objectively measurable environments.

Conclusion

Nex-N1 presents a comprehensive solution to the data and environment bottlenecks in agentic model training. Its unified, modular ecosystem—comprising NexAU, NexA4A, and NexGAP—enables unprecedented scaling and realism in agent-environment interaction. Empirical results establish Nex-N1 as a leading open-source agentic model, with versatile, robust capabilities across established benchmarks and practical domains. The open-sourcing of the ecosystem and its data is positioned to accelerate research towards generalized, autonomous agentic intelligence.