Moonshot AI’s July 2026 launch of Kimi K3, a 2.8-trillion-parameter MoE model with native vision and a 1M-token context window, established the current Kimi frontier on Humanity’s Last Exam at roughly 43.5–46.9% text-only and 56% with tools, placing it behind leading Claude and GPT variants. Trader sentiment reflects Kimi’s open-weights release strategy, which enables community fine-tuning and agentic tool-use improvements through year-end, alongside incremental updates like the September Kimi K2.8 Preview that approaches K3 performance on coding and reasoning tasks. HLE’s 2,500 expert-level questions across math, science, and humanities create a high bar for rapid gains, yet Moonshot’s emphasis on long-horizon agents and large-scale context supports continued optimization before December 31. Key catalysts include any new Kimi variants or verified benchmark submissions that could push scores toward 50%+ or higher.