A single prompt to a local Qwen3 28B quant produced a fully playable, procedurally generated Three.js desert flight demo in one HTML file.
Adapted from @Oluwaphilemon1Qwen3.8-27B-GSQ-RCO-IQ3_S keeps surprising me on a 3090. This time AJ gave it a much more demanding prompt: Build a procedurally generated endless desert world that I can fly through. No multi-agent setup. No back-and-forth prompting. No hand-written Three.js code from me. Just a one-shot prompt and a single HTML file. And it actually delivered a playable world. The scene generates the desert procedurally, giving you an environment that feels much larger than what is actually being rendered at once. That kind of task is a pretty good stress test for a local coding model because there are several things it has to get right at the same time. It needs to understand the visual goal. It needs to structure the Three.js scene. It needs procedural generation rather than simply placing a few static objects. It needs camera and flight controls. It needs to keep everything self-contained in one HTML file. And most importantly, the result has to actually run. This is why I’m starting to care less about whether a local model looks impressive on a benchmark and more about what it can produce from a blank file. Qwen3.8-27B running locally on a 3090, in this quant, is already getting into territory where you can throw fairly ambitious browser experiments at it and get something usable back. A few years ago, I’d expect to iterate heavily on something like this. Now the first attempt can already be the interesting part. And this is only one result. I’m testing more prompts with this quant to see where it starts breaking down. Local AI is getting much more interesting when the output isn’t just text, but an actual working thing you can open and play with. https://t.co/7Hjd0lwt8e