News
Hexagon and 12 balls: a model of over 80 GB on an RTX 3060, two versions side by side
Qwen3.8-Flash-Next IQ3_S on an RTX 3060 12 GB with only 24 GB of single-channel RAM. One prompt, two versions of the result, both running live below with the code unedited.
This is a small part of our larger visbench experiment (contour series): the local model gets one visual programming task, and the result is checked in a browser.
Configuration
- GPU: RTX 3060 12 GB.
- CPU: Intel Core Ultra 7 265.
- RAM: only 24 GB as a single module, i.e. single-channel (half the speed that dual-channel memory allows).
- Model: Qwen3.8-Flash-Next IQ3_S on the Strata 0.1.52 engine.
The contrast: the model itself takes more than 80 GB, while the machine has only 24 GB of RAM in a single module. Only 22 GB of that is usable, because the BIOS reserves 2 GB for the graphics built into the motherboard, and Ubuntu 24.04 Desktop itself also uses part of the memory.
Side by side
Both files are the model's originals, unedited (no external calls), running right in your browser.
The prompt (verbatim)
Create a complete, standalone HTML file containing an animated 2D physics simulation.
Requirements:
• A large regular hexagon rotates continuously at a constant angular velocity.
• Exactly 12 colorful balls move inside the hexagon.
• The balls are affected by gravity.
• Balls bounce realistically off the rotating walls, taking wall velocity into account.
• Balls must not escape the hexagon.
• Add damping and restitution.
• Display smooth animation using HTML Canvas and JavaScript.
• Use a dark background and attractive neon colors.
• The simulation should fit a 1280×720 browser window.
• Show a small FPS counter.
Restrictions:
• Use only HTML, CSS and vanilla JavaScript.
• No external libraries, CDN, network access or image assets.
• Return the complete source code of a single index.html file.
• Do not include explanations or Markdown code fences.
• Do not ask questions.
• Do not use external tools.
The application must run immediately when index.html is opened in a browser.
Version 1: the first run
One request, and the code was not repaired (only the Markdown fence around the answer was stripped). 83.82 s in total, first token after 5.13 s, 57.0 tok/s, 4489 tokens, 11901 bytes, no external calls. Check in an isolated browser: PASS, 12 balls, 0 escapes, 0 console errors. The code contains a hard clamp that holds the balls inside.
The first “brain fix”
This is not fine-tuning: the model itself was not changed. We changed only the wrapper around it: error control, a check of the result in an isolated browser, verdicts and up to two attempts. With the same prompt, the model produced a new version.
Result after the wrapper
With the same prompt, version 2 passed 5 of 5 seeds on the first request, and the code was not edited.
Seed 0, shown below: 69.1 s in total, first token after 9.9 s, 54.4 tok/s, 3225 tokens, 0 JavaScript errors.
Download all 5 versions after the wrapper, with screenshots (zip, 1.0 MB)