A 35-billion-parameter agent that punches like a trillion-parameter model

Short summary

Shanghai AI Lab's 35B-parameter Agents-A1 model matches trillion-parameter models on complex, long-horizon agent tasks by training on longer action sequences (averaging 45K words per task) rather than scaling raw parameters. The approach combines distillation and mixture-of-experts architecture to keep computational costs manageable. Implication: effective agents may depend more on superior training data and intelligent architecture than brute-force parameter scaling.

• •35B-parameter model matches trillion-parameter models on agent tasks through longer training sequences, not scale
• •Uses distillation and mixture-of-experts for efficiency
• •Suggests agents need better training data and architecture more than raw parameter count

Generated with AI, which can make mistakes.