Moonshot's K3 matches GPT-5.6 benchmarks but fails at math
moonshot ai released open weights for its kimi k3 model, which delivers 2.5x intelligence per compute unit but struggles with math and cyber tasks despite competitive benchmark scores.
moonshot ai released the model weights and technical report for kimi k3 on hugging face, alongside high-performance attention kernels and tools for running agents at scale. the company claims the architecture delivers 2.5 times more intelligence per unit of compute and competes with models like fable 5 and gpt-5.6 sol on general benchmarks. however, the uk cyber institute found that k3 lags behind frontier models in cyber operations and mathematics. these performance gaps suggest the model may rely on distillation techniques, a method gaining legitimacy among open-weight advocates despite previous criticism.