The scaling of executable agent training data is often bottlenecked by substrate-first methods, which tie task generation to predefined tools, repositories, or skill graphs. Expanding coverage typically requires manual infrastructure expansion, and each new domain demands a bespoke pipeline, leading to task distributions that reflect substrate convenience rather than real-world demand.
We introduce NexForge, a requirement-first framework that compiles free-form capability requirements into executable agent training data.
NexForge first conducts research-based demand discovery to identify representative task forms, realistic scenarios, and their prevalence. It then applies distribution-aware task compilation, automatically retrieving or constructing necessary files, repositories, dependencies, and runtime configurations for each task, followed by teacher rollout collection and trajectory distillation.
The same pipeline, without any domain-specific infrastructure, produces 3,600 terminal tasks and 2,000 office tasks, improving Qwen3.5-35B-A3B Base from 22.5% to 52.0% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval; scaling to 43.2K terminal tasks reaches 58.4%, surpassing Claude Opus 4.6.
Further scaling, NexForge-synthesized data contributes to the training of Nex-N2, a family of publicly available agent models that elevate Qwen3.5-35B-A3B to 75.3% on Terminal-Bench 2.1 and to 1585 Elo on GDPval -- achieving state-of-the-art open-source performance and surpassing several frontier proprietary systems. Nex-N2 models are available at Nex website.
Blogger's Review: NexForge's requirement-first synthesis method effectively addresses the limitations of traditional substrate-based approaches by generating high-quality training data through automation and distribution-aware strategies. This innovation not only enhances model performance but also lays a foundation for the development of open-source agent models, showcasing significant practical value and forward-thinking potential.