tl;dr: Moonshot has released Kimi K3, a 2.8T-parameter open-weight model with Kimi Delta Attention, Attention Residuals, and a 1M-token context window, activating 16 of 896 MoE experts. It claims frontier-level performance on coding, agentic, and reasoning benchmarks—outperforming GLM-5.2 and Claude Opus 4.8 in many cases, though still trailing Claude Fable 5 and GPT 5.6 Sol. Available now via API ($3/MTok input, $15/MTok output), with full weights slated for release by July 27, 2026.