Kimi K3: Open Frontier Intelligence

The paper introduces Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion activated parameters per token, native vision, and a 1-million-token context window. Its architecture combines Kimi Delta Attention, Attention Residuals, and Stable LatentMoE, which routes each token through 16 of 896 experts. Post-training uses reinforcement learning across general, agentic, and coding tasks with multiple reasoning-effort levels. The authors report roughly 2.5x better overall scaling efficiency than Kimi K2 and claim frontier-level results across coding, agents, knowledge, reasoning, and vision. Full model weights are released, although the abstract says performance still trails leading proprietary systems.