A small AI model outperforms giants by learning to explore more effectively.

Researchers have developed a new, compact 4-billion-parameter AI agent that rivals much larger models. They overcame three key training problems: forgetting, noisy feedback, and reasoning errors. The resulting model, AgentCPM-Explore, matches or beats 8-billion-parameter models and outperforms giants like Claude-4.5-Sonnet in five benchmarks. It achieved 97% accuracy on a complex reasoning test. This proves small models have untapped potential, limited by training stability, not size.