Composite-Bench released: long-horizon computer-use benchmark, GLM-5.2 beats Kimi K3 by 32 points, surpasses all closed models except Claude

Today sees the release of Composite-Bench, a long-horizon computer-use benchmark with certified-optimal answers and verified compute. The strongest open-weights model is not Kimi K3; GLM-5.2 beats it by 32 points and outperforms every closed model tested except Claude.

• benchmark
• GLM 5.2
• Computer Use
• Kimi K3
• Composite-Bench