Introducing Agents-A1: A 35B Agentic Model Achieving Trillion-Parameter Performance

Researchers have unveiled Agents-A1, a 35 billion parameter Mixture-of-Experts model that achieves performance levels comparable to trillion-parameter models by focusing on agent horizon scaling. The study explores two main aspects: extending long-horizon trajectories and enhancing the abilities of heterogeneous agents. The team developed a long-horizon knowledge-action infrastructure that integrates external knowledge with actions and observations, producing agentic trajectories averaging 45,000 tokens. Agents-A1 was trained using a three-stage approach, including full-domain supervised fine-tuning, domain-level teacher models for specialized expertise, and a novel multi-teacher domain-routed on-policy distillation method. The model has demonstrated superior performance on several long-horizon agent benchmarks, outperforming notable models like Kimi-K2.6 and DeepSeek-V4-pro across various metrics, and aims to offer a practical solution for scaling agent performance.