Lab Log / 2026-07-07 / Local AI

Seven models, one prompt,
three Earths.

We handed the same one-line prompt to seven AI models — four running locally on a single desktop, three in the cloud — and asked each to build a rotating, textured 3D Earth. Only three produced a working render. Two of them ran on the box on our desk, for nothing.

All seven renders, running live — same prompt, same judge 3 rendered · 4 broke
Top row — LOCAL, on the desktop
Ornithlocal · works
Chadrocklocal · works
Gemmalocal · broke
Qwenlocal · broke
Bottom row — CLOUD · free NVIDIA API + Canopy Wave
Kimi K2.6NVIDIA (free) · broke
MiniMax M3NVIDIA (free) · works
GLM 5.1Canopy Wave · broke
The test

The prompt — identical to all seven

"Build a realistic 3D Earth that rotates continuously — Three.js, textured sphere, sunlight, a starfield, standalone in a browser."

The judge & the rig

  • Pass = a real rotating Earth renders
  • Fail = broken, blank, or wrong
  • Local box: one 128 GB desktop
  • Models at ~4.3 bits/weight
The board
ModelWhereTimeResult
1Ornith 35Blocal10.0sworks — cleanest of all seven
2Chadrock 35Blocal13.0sworks — custom atmosphere shader + clouds
3MiniMax M3cloud · nvidia237sworks — best-looking, but ~4 minutes
Gemma 26Blocal8.8sbroke — invented an npm package (404)
Qwen 30Blocal14.7sbroke — missing one config line → blank
Kimi K2.6cloud · nvidia25.4sbroke — one garbled line → syntax error
GLM 5.1cloud · canopy wavebroke — wrong texture in the cloud slot

Local coding models went 2-for-2. Cloud went 1-for-3.

Local models ran on one desktop via llama.cpp. Cloud models were served on the free NVIDIA API (Kimi K2.6, MiniMax M3) and Canopy Wave (GLM 5.1) — hosted, not local. Even the free cloud tier didn't change the outcome.

The two winners

Prettiest vs. worth it

It wasn't a clean sweep, and pretending it was would be a lie. Here's the honest split:

Cloud · best-looking

MiniMax M3

237 seconds · free NVIDIA API

The only render with real weather — a proper cloud layer, atmospheric glow, a soft day/night terminator. It genuinely looks like a photo from orbit. It also thought for four minutes on someone else's machine to get there.

Local · best value

Ornith 35B

10 seconds · on our desk · $0

A correct, clean, photorealistic Earth — same base texture, ocean highlight, atmosphere glow, starfield. Written in 945 tokens, ten seconds, no bill, no data leaving the building.

The entire visual gap between them was one cloud layer. Same base Earth, same continents, same glow. Strip MiniMax's clouds and they're near-twins — and that single layer of extra polish cost 24× the time and a rented datacenter.

What it means

Capability is table stakes.
Reliability is the moat.

Not one of the seven failed on the actual 3D graphics. Every model knew the concept cold — scene, lighting, sphere, animation loop, all correct. The four that broke tripped on a detail: a made-up dependency, a missing line, one stuttered token, a mis-wired asset. The winners weren't smarter about graphics. They were more careful.

And the sharper point is the price tag on polish. The best render came from the cloud — but a 4-bit model on one desktop got 90% of the way there instantly, for free, with nothing leaving the room. The question isn't whether local beats the cloud. It's this: when local is good enough, instant, and free, the cloud has to justify its bill on every single query.

Receipts
01

7 models, 1 identical prompt, 3 clean renders.

02

Best value: Ornith, local — a correct Earth in 10 seconds for $0.

03

The cloud's prettier render took ~4 minutes — 24× the time — for one extra cloud layer.

04

Cloud models went 1-for-3. Local coding models went 2-for-2.

05

No model failed on the graphics — every failure was a small precision slip.

06

Hardware: one 128 GB desktop, models running at ~4.3 bits per weight.

The fine print, because receipts beat hype: one task, graphics only — don't read it as "model X beats model Y at everything." One cloud model (GLM) ran through an agent harness rather than a direct prompt, so its conditions differed. And all four failures are fixable one-liners — these models are one token from working, not incapable. The story isn't that the cloud is bad. It's that local is now genuinely in the fight.

← Back to Alpha Ranch