Skip to content

kimi k3 vs glm 5.2

pulse Three retro CRT monitors labeled kimi k3 and glm 5.2 showing a pixel game, bright fish tank, and a dark near-empty tank.

@kimi_moonshot k3 just knocked @zai_org's glm 5.2 out of the top 3 on code @arena | webdev:

• #1 kimi-k3 – 1679 (debuts straight at the top, preliminary)
• #2 claude-fable-5 – 1631
• #3 gpt-5.6-sol-xhigh – 1618
• #4 glm-5.2 (max) – 1587

glm was sitting in the top 3. kimi walked in at #1 and shoved it down to 4th. so we put the two of them head to head

our test – 3 prompts, single-file html, @threejs, fully procedural, no assets, orbitcontrols, pcfsoft shadows, acesfilmic + srgb:

• late-80s living room – an original nes wired to a crt that is actively playing a period-correct pixel game, drawn live to a canvastexture, plus a working tv power toggle

• vintage steel road bike on a kickstand in a workshop – real drivetrain kinematics: turn the pedals, the cranks drive the chainring, the chain drives the rear sprocket, both wheels spin at the correct gear ratio

• glass aquarium in a quiet study – 8–15 boids fish of several species with tail undulation, transmissive glass, caustics, bubble streams, a feed button

we ran it on @aimlapi

results:

- cost:
#1 kimi k3 – $1.02
#2 glm 5.2 – $1.57

- tokens:
#1 kimi k3 – 193,632
#2 glm 5.2 – 286,263

- lines of code:
#1 glm 5.2 – 3,762
#2 kimi k3 – 3,114

- generation time:
#1 glm 5.2 – 64.0 min
#2 kimi k3 – 70.6 min

observations:

• kimi is cheaper and lighter – 35% less money, 32% fewer tokens – for three working scenes. per 100 shipped lines it's $0.033 vs glm's $0.042

• glm reasons roughly 3x harder than kimi (171k vs 66k reasoning chars on the nes alone). you're paying for a longer argument with itself, not a bigger answer

• kimi handles light better. every one of its rooms is bright, warm and readable. glm renders everything in gloom: a "saturday evening" nes room you have to squint at, an aquarium so dark you lose the fish in it. same prompt, same tone-mapping instruction, opposite result

the scenes:

• nes – tie, both nailed it. kimi warm and cozy, glm moody and dark, both fully working

• bike – kimi wins. its only flaw: no crank arms, so the pedals spin in the air. glm broke more: handlebars fused to the front wheel, window perpendicular to the wall, pedals frozen on the cranks

• aquarium – kimi in a landslide. bright, ~10 fish schooling, runs smooth (minor nan console warnings). glm's is near-black murk you can't find the fish in, and it lags at low fps

a note in fairness to both: nobody nailed the bike drivetrain. the pedals, cranks and wheels linkage is the single hardest mechanical ask in the set, and each model fudged it its own way: kimi dropped the cranks, glm froze the pedals.

kimi k3 is the stronger model here. it's cheaper, lighter on tokens, lights a room properly, and ships fewer and smaller mistakes across all three scenes. it earns that #1 spot

0:00
/0:51

Stay in the loop

Get the latest AI news delivered to your inbox weekly

Thanks for subscribing!