Shanghai AI Lab's 35B-parameter model rivals trillion-parameter AI on agent tasks

Shanghai AI Lab has developed Agents-A1, a 35-billion-parameter AI model that reportedly matches the performance of trillion-parameter models on complex, multi-step agent tasks, according to a paper published on arXiv on June 30, 2026. Instead of increasing parameter count, the researchers scaled the 'horizon' — training the model on much longer and more varied action sequences averaging around 45,000 words per task. The team also used a distillation approach, training specialist teacher models in individual domains and merging their expertise into a single student model, while a mixture-of-experts architecture keeps computational costs manageable. Agents-A1 claims competitive results on benchmarks involving tool use, web browsing, and scientific reasoning against far larger systems. The findings are self-reported and have not yet been independently verified, meaning real-world reliability beyond curated benchmarks remains to be established.