tl;dr: Moonshot's Kimi K3 is a 2.8T-parameter open model with native vision, 1M-token context, and a novel architecture (Kimi Delta Attention + Attention Residuals, 16-of-896 MoE sparsity) claiming 2.5× better scaling efficiency than K2. It reportedly trails Claude Fable 5 and GPT 5.6 Sol overall but competes strongly on long-horizon coding, agentic tasks, and multimodal benchmarks, with demos including autonomous chip design, a from-scratch Triton-like compiler, and video editing. Available now via API ($0.30/$3/$15 per MTok) with full weights to be released July 27, 2026.