4 chinese models, 1 animation task: glm spent 21 minutes and won

glm 5.2 max vs kimi k2.7 code vs minimax m3 vs mimo v2.5 pro: making the moon go around the earth

the top-3 open weight labs – @Kimi_Moonshot , @MiniMax_AI and @Zai_org – all dropped a new model in june

so we put them head to head on arena ai with one simple task: generate an animation of the earth and moon in space

we also threw in @XiaomiMiMo v2.5 pro as a cheaper option to see if it'd fall behind

biggest surprise: glm takes ~20x longer to produce the same volume of code

it's not a bug. the model just reviews everything thoroughly before committing

time spent:
• minimax m3 – 47 sec
• kimi k2.7 code – 51 sec
• mimo v2.5 pro – 1 min 27 sec
• glm 5.2 max – 21 min 1 sec (!)

lines of code written:
• kimi k2.7 code – 297
• minimax m3 – 384
• glm 5.2 max – 462
• mimo v2.5 pro – 503

but line count tracks ambition, not quality. so we actually read all four

glm: the most serious one. only model that built genuinely seamless procedural textures (no visible seam on the sphere), put clouds on a separate rotating layer, wrote a real fresnel atmosphere shader and got the full color pipeline right. that 21 min bought something

mimo: richest visuals. hand-drawn recognizable continents, separate bump + spec maps, multi-layer sun glow, proper per-star twinkle. loses to glm only on a texture seam and some fake fixed-direction lighting. great value for a "cheaper" model

minimax: cleanest, most readable code and runs great. but every star twinkles in unison (it even computes per-star data and never uses it), and the atmosphere is a flat sprite. solid, just cuts the two corners that matter most for the wow

kimi: most compact. nice ideas – real per-star shader, even attempts shadow mapping. but the earth is the least detailed and it pulls in the entire tailwind cdn to basically do nothing

next up we're hoping to throw a new @deepseek_ai into the ring – fingers crossed it drops soon