Moonshot AI’s July 2026 launch of Kimi K3, a 2.8-trillion-parameter mixture-of-experts model, has been the main catalyst lifting Kimi scores on Humanity’s Last Exam. The release delivered strong gains in tool-augmented reasoning and agentic workflows, placing K3 in the mid-50s percent range on recent leaderboards and ahead of most open-weight rivals. Earlier K2-series variants had already shown notable tool-use improvements, with scores rising from the low 30s to over 44 percent when search and code interpreters were enabled. Anthropic’s Claude Fable and Opus variants continue to lead overall HLE performance, but Moonshot’s rapid iteration and focus on long-context agent capabilities keep Kimi competitive. Traders are watching for any late-2026 model updates or inference optimizations that could further close the gap before year-end.