Agents-A1: 35B-Parameter MoE Agentic Model Matches Trillion-Parameter Performance via Long-Horizon Training, Tops HF Daily Papers With 34 Upvotes

A new paper from Intern Science introduces Agents-A1, a 35B Mixture-of-Experts agentic model trained on a Knowledge-Action Graph corpus of ~100K trajectories averaging 45K tokens each. The model reportedly outperforms 1T-parameter models (Kimi-K2.6, DeepSeek-V4-Pro, GPT-5.5) on multiple long-horizon agent benchmarks including GAIA (96.0%), SEAL-0 (56.4%), and IFBench (80.6%), arguing that scaling interaction horizon and infrastructure beats parameter scaling. Limitations remain on sustained ML engineering tasks where it lags GPT-5.5.