Kimi K3 ranks third on new DeepRSI benchmark, top Chinese model

DeepRSI shifts the evaluation target from static task performance to recursive self-improvement and autonomous research, a capability set that matters more for long-horizon R&D than for chat benchmarks. Kimi K3 landing third, ahead of several US frontier models and atop Chinese peers, suggests the gap at this specific capability is narrower than headline model rankings imply. The open question is whether DeepRSI's methodology holds up to scrutiny and whether other labs reproduce these placements, since a single first-run leaderboard is a weak signal on its own.