35B Model Matches Trillion-Parameter Performance By Scaling Task Complexity

For years, the AI industry equated scale with capability. Shanghai AI Laboratory has published a paper challenging that assumption.

Their model, Agents-A1, operates on a 35-billion-parameter Mixture-of-Experts architecture. It matches, and sometimes outpaces, systems thirty times its size. The key innovation was not expanding the model’s parameters, but training it on longer, more complex task sequences.

The model undergoes a three-stage training protocol. It begins with broad, supervised fine-tuning. It then learns from domain-level teacher models. A final on-policy distillation stage allows it to absorb knowledge from multiple specialized sources simultaneously.

The benchmark results highlight this shift in strategy. Agents-A1 scored 56.4 on the complex agent benchmark SEAL-0, 80.6 on the instruction-following test IFBench, and 96.0 on the general assistant evaluation GAIA. The system also supports tool usage and function calling, enabling real-world interaction with external software and APIs.

Developed by InternScience at the Shanghai AI Laboratory, Agents-A1 was open-sourced on June 30, 2026, under an Apache-2.0 license. The research argues that extending a task’s horizon offers a powerful second scaling axis beyond raw parameter count.