Open weight Kimi K3 crushed cloud frontier GPT 5.6 at 3D destruction physics! Moonshot AI became the first lab to open-source frontier model weights. We tested them on 8x B300 We gave four models the same task: build three self-contained HTML scenes with real physics Prompts: – A monster... truck crushing a row of cars – Two cars jumping a canyon and colliding head-on mid-air – A giant anvil drop test flattening cars one by one Outputs: – Kimi K3 (local): 32.7K tokens, $0 – GPT 5.6: 19.2K tokens, $0.30 – Grok 4.5: 55.2K tokens, $0.45 – GLM 5.2: 66.8K tokens, $0.15 Kimi K3 made the most realistic scenes and won all three against every other model. Its truck rides over the cars and crushes them one by one, and the wrecks stay crushed. The canyon scene shows the gap best: Kimi's cars smash into each other and fall into the gorge, GPT's cars fly past the crash and land with no damage. Grok 4.5 and GLM 5.2 surprised us with the collision animation in the second scene: the impact moment is the most epic of the whole run, bumpers fold, wheels fly off, debris flies everywhere. Their other scenes are nowhere near that levelshow more

run Kimi K3 locally very soon here:

cooking smth 👨‍🍳😤

What’s up with grok

bullish on kimi

Open weight models win!

What was the verdict on qwen 3.8?

$0 is not useful information, would be better to show the rent price to run it for the time needed for the test. Here is an estimate from Grok:

That is an absolute masterclass by Kimi K3! 🤯.

insane it topped hf trending with 4k likes in 30 mins

@atomicbot_ai How can I use oss Kimi on my Mac 48gb?

The way GPT’s cars are flying past each other looks more like tunneling than bad physics. wondering, if any of them pull in a physics lib or was it hand-rolled integrators.

@grok tu nous a fait quoi là. Les roues sont meme pas dans le bon sens.

insane it topped hf trending with 4k likes in 30 mins

Reproduction as the product is the right frame. Re-running a leaderboard's own 98,088 public battles put its #1 at rank CI [1,1] — a certification, and worth more than a debunk. Has anyone actually re-run the Frontend Code Arena ranking end to end yet?

Everyone will read this as "open beat cloud." The sharper story is the token count. Kimi did it in 8,448. Grok and GLM burned 18k-19k for a worse result. Open weight AND leaner. That's the part that should scare the labs.

Kimi K3 local physics hits different the truck rides over and leaves real wrecks while GPT just yeets past clean. Open weight on B300s at zero cost is the real shift.

At some point AI will be better: still a ways to go. Here's a physics simulator I wrote ~20 years ago and recently updated for 2026 (does physics directly on bicubic curved surfaces for the tracks, not triangles).

the Kimi physics demo makes open weights feel real. normal people understand flying debris faster than they understand another leaderboard table

kimi real G underdoge

These comparisons keep getting more interesting. The gap between closed and open models is closing faster than expected in certain tasks.

Great perspective — we cover this kind of AI development daily at @@ainewsusa

3d scene demos reward visual luck did you run repeatable physics tests too

kimi used twice the tokens and still won which is either impressive or a different kind of problem

The real shift here isn't the Frontend Code Arena ranking, it's that 8x B300 can reproduce it. A year ago 'open weights beat frontier' was a tweet; now it's a bill of materials anyone with a credit line can order. Moonshot just turned a benchmark into infrastructure.

i was also testing it! This model is a gem!

this is the price in the UK to run the full model

这类场景对物理细节还挺刁钻的

Open-weight Kimi K3 beating a frontier cloud model on this kind of physics task is notable. The gap is closing faster in certain domains than many expected.

👀 impressive adding it to the list