Moonshot AI’s Kimi K3 is an open frontier model that sits close to the best closed models on most benchmarks, and its forty-seven technical report reveals parts other labs keep private. The architecture is a stack of small changes: K3 is a 2.8-trillion-parameter mixture-of-experts model. It uses a fixed-size running state instead of a growing cache, making a million-token context affordable; most layers use this cheap version, and every fourth does full attention. Positions aren’t encoded explicitly, so recurrence carries them, letting the model stretch to a million tokens without rescaling hacks. Each layer can look back at every layer beneath it, and each token is routed to 16 of 896 experts, sparser than before. Moonshot reports a 2.5× gain in scaling efficiency over Kimi K2, roughly the same quality for under half the training compute, crediting the architecture together with refined data and training recipes without apportioning the gain. The attention swaps a growing filing-cabinet cache for a fixed-size state, avoiding cost that climbs with the square of length. Each piece is an ordinary, well-tested idea rather than a single breakthrough, and the decisions that made it good—training, reinforcement learning, serving cost—stay in-house, though Moonshot’s report walks through parts other labs keep private.