Home
How to Scale Large Models: Yang Zhilin's GTC Strategy on Token Efficiency and Agent Clusters

The ticket to the second half of the large model era is no longer about simply scaling compute, but a fundamental rethinking of the underlying architecture.
At the NVIDIA GTC 2026 conference on March 18, Moonshot AI founder Yang Zhilin delivered a highly anticipated keynote. This marked his first comprehensive public outline of the core technical roadmap behind the Kimi K2.5 model, providing fresh perspective on large model evolution in the "post-scaling" era.
Yang Zhilin stated that to break through current intelligence limits, a complete restructuring of key technologies like optimizers, attention mechanisms, and residual connections is essential. He framed Kimi's evolution across three synergistic dimensions:
Token Efficiency: Eliminating resource waste to pursue an even more extreme compute-to-performance ratio.
Long Context: Continuously deepening Kimi's long-context memory advantage to process information at a massive scale.
Agent Cluster: Intelligence is evolving from individual agents to dynamically generated "digital clusters."
In Yang Zhilin's view, scaling has now evolved into finding scale effects in efficiency, memory, and automated collaboration. Multiplying the gains from these three dimensions could unlock intelligence levels far beyond current capabilities.
According to earlier announcements, the Kimi K2.5 model launched in early January already demonstrates this "all-around" capability. As Moonshot AI's most powerful open-source model to date, it features a native multimodal architecture, achieves state-of-the-art (SOTA) performance in code and visual understanding, and supports flexible switching between "thinking" and "non-thinking" modes to precisely adapt to agent-based tasks.
As Moonshot AI's technological approach becomes clearer, the large model competition is shifting focus from "parameter count" to "intelligence density." With agent clusters emerging as a potential ultimate form of future intelligence, whether Kimi can achieve a breakthrough under Yang Zhilin's "three-dimensional multiplication" framework has become a key industry focus.
Related article
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber
Related Special Topic Recommendations
Comments (0)
0/500

The ticket to the second half of the large model era is no longer about simply scaling compute, but a fundamental rethinking of the underlying architecture.
At the NVIDIA GTC 2026 conference on March 18, Moonshot AI founder Yang Zhilin delivered a highly anticipated keynote. This marked his first comprehensive public outline of the core technical roadmap behind the
Yang Zhilin stated that to break through current intelligence limits, a complete restructuring of key technologies like optimizers, attention mechanisms, and residual connections is essential. He framed Kimi's evolution across three synergistic dimensions:
Token Efficiency: Eliminating resource waste to pursue an even more extreme compute-to-performance ratio.
Long Context: Continuously deepening Kimi's long-context memory advantage to process information at a massive scale.
Agent Cluster: Intelligence is evolving from individual agents to dynamically generated "digital clusters."
In Yang Zhilin's view, scaling has now evolved into finding scale effects in efficiency, memory, and automated collaboration. Multiplying the gains from these three dimensions could unlock intelligence levels far beyond current capabilities.
According to earlier announcements, the Kimi K2.5 model launched in early January already demonstrates this "all-around" capability. As Moonshot AI's most powerful open-source model to date, it features a native multimodal architecture, achieves state-of-the-art (SOTA) performance in code and visual understanding, and supports flexible switching between "thinking" and "non-thinking" modes to precisely adapt to agent-based tasks.
As Moonshot AI's technological approach becomes clearer, the large model competition is shifting focus from "parameter count" to "intelligence density." With agent clusters emerging as a potential ultimate form of future intelligence, whether
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber











