Kimi K3.1 Pro, Moonshot AI’s late-October 2026 release, currently posts the highest reported Humanity’s Last Exam scores at 70% in tool-augmented settings, ahead of Claude Opus 5.5 (67.7%) and other frontier models. The 2.8-trillion-parameter MoE model with 1M-token context was post-trained on expanded reinforcement learning across coding, agentic, and knowledge domains, producing measurable gains on Terminal-Bench and long-horizon tasks while deliberately limiting cyber-related capabilities. Different HLE variants—text-only, HLE-Diamond without tools, or verified subsets—yield lower scores for the same model (around 52% or 29.5%), illustrating protocol sensitivity. Full open weights are scheduled for release by November 4, which could trigger independent re-evaluations and further shift trader views on whether Kimi holds the 2026 peak.
Експериментальне резюме, згенероване ШІ з посиланням на дані Polymarket. Це не торгова порада і не впливає на вирішення цього ринку. · Оновлено1H6H1D7D30DALL
View resolved

Обережно з зовнішніми посиланнями.
Обережно з зовнішніми посиланнями.
Часті запитання