Local Kimi K3 beats GPT 5.6 in physics crash simulations

In a test on atomic.chat, locally hosted Kimi K3 (8X B300s) produced the most convincing physics in three browser-based crash scenes, outperforming GPT 5.6. Kimi K3 used 32.7K tokens at $0 cost, while others cost $0.15-$0.45. The test involved generating self-contained HTML simulations; Kimi K3 handled damage persistence best.

• 本地模型
• 物理模拟
• Kimi K3
• 比较测试