Yang Zhilin disclosed Kimi's technical roadmap at GTC: focusing on Token efficiency, long context, and Agent clusters

At NVIDIA’s GTC conference in 2026, Moonshot’s Kimi founder Yang Zhilin delivered a public keynote speech. He said that to continuously push past the ceiling of intelligence in large models, it is necessary to redesign the underlying foundations such as optimizers, attention mechanisms, and residual connections. After Kimi K2.5 was officially released at the end of January this year, Yang Zhilin, for the first time in his talk, systematically disclosed the technical roadmap behind the model. He summarized Kimi’s evolutionary logic as three dimensions of resonance: Token efficiency, long context, and Agent Swarms. “Scaling is no longer just about piling up resources; instead, we need to seek scale benefits simultaneously in computational efficiency, long-range memory, and automated collaboration. If we can multiply the technical gains across these three dimensions, the model will exhibit an intelligence level far beyond the current state.” In addition, he judged that future forms of intelligence will evolve from a single agent to dynamically generated clusters. (China’s Science and Technology Innovation Board Daily)