Moonshot AI releases Kimi K3, a 2.8T parameter open-weights model

The 2.8T parameter MoE model features native vision support and a 1M token context window, delivering 2.5x better scaling efficiency than Kimi K2. It achieves top-tier performance on benchmarks like DeepSWE, where it outperforms GPT-5.6 Sol on pass@4 metrics while costing 64% less per task.