Free 32 GB Kaggle GPU vs my 12 GB RTX 3060
The same model, Qwen3.8 27B, on the same two prompts. On Kaggle's free T4 x2 with a 4-bit file of 17 GB, about 12 tokens per second. On my RTX 3060 with a tighter 3-bit file of 9.7 GB, about 19 tokens per second. The cloud builds work after small marked fixes; the two from the smaller file do not work. Two builds are a sample, not a benchmark. Everything runs in your browser.
The builds
A wrong bracket, a missing Seq.play() and a variable that overwrote start(): all marked in the file. Then it plays, and Generate song works.
Exactly what the model wrote: the script stops on its first line of bad syntax.
A const and a function with the same name, and a function called window() that overwrote the browser's window. Then the 60 seconds play through all the life stages.
Exactly what the model wrote: the script does not parse.
The interface draws, but the script has a syntax error. One marked fix below shows a second bug behind it. Published as written.
The first bug fixed and marked: pressing play stops on a second bug in the audio code.
It reached my limit of 24000 tokens before finishing the file, and it has a syntax error too. Published as written.
Time per build
| Where | Build | Time | Tokens | Tokens/s | Result |
|---|---|---|---|---|---|
| Kaggle free GPU · 2 x T4 | Nocturne (3 fixes) | 32.8 min | 24000 | 12.2 | Works after fixes |
| Kaggle free GPU · 2 x T4 | Every Frame (2 fixes) | 32.1 min | 23500 | 12.2 | Works after fixes |
| My RTX 3060 · 12 GB | Signal, as written | 13.8 min | 16296 | 19.7 | Does not run |
| My RTX 3060 · 12 GB | Every Frame, cut off | 20.9 min | 24000 | 19.2 | Does not run |
The exact prompts are for Ruocco Academy members. The Academy is not open yet: join the waiting list and you get one email when it opens. The demos on this page are free to use for everyone.
Join the waiting list