kimi-k2.7-code vs minimax m3 on a frontend task

new kimi-k2.7-code by @Kimi_Moonshot shipped today

we decided to compare it with @MiniMax_AI m3 on @arena ai – fresh rival, shipped recently, and outscored kimi k2.6 on the artificial analysis intelligence index by just one point. close enough to make the head-to-head interesting

asked to make an animation of a tropical waterfall

• minimax spent 4 min 35 sec
• kimi spent 13 min 16 sec

• minimax wrote 1,634 lines
• kimi wrote 1,120 lines

minimax packed more into the scene – extra flower species (orchids, hibiscus, heliconia), atmospheric mist, floating dust motes, numbered drawing layers for readability. but it stuffed all 1,400 lines into a single function, hardcoded the canvas to 1920x1080 so it'll look blurry on retina/4k screens, and tied animation speed directly to the browser's frame timer (so it'll run faster or slower depending on your device)

kimi has fewer scene elements but separated the code into 10 named data structures at the top, detects your screen's pixel density and scales the canvas to match (sharp on any display), uses delta-time so animation runs at the same speed everywhere, and built the audio with a proper master volume bus so muting actually works cleanly

verdict: minimax shipped more pixels faster, kimi shipped cleaner foundations slower. different bets on what "done" means