Imagine creating a 2D Mario-style game in just a few seconds. Sounds like a fantasy, right? But what if I told you it’s now possible with Qwen3-Coder, a newly released open-weight model from Alibaba for code generation?
Recently, I came across an experiment where Qwen3-Coder was connected to Cursor IDE via a standard OpenAI-compatible API. The goal was to give it a high-level task: create a 2D game like Super Mario. And you know what? It worked like a charm.
The model asked if assets were present in the folder, installed pygame, and added a requirements.txt file. It even generated a clean folder layout with main.py, a README, and placeholders. But that’s not all – it also implemented player physics, coins, enemies, collisions, and a win screen. And the best part? The code worked directly, with no edits required.
What makes this interesting is that the model handled the full task lifecycle from a single prompt, with no hallucinated dependencies or syntax errors. The inference cost was around $2 per million tokens, which is impressive. It’s also worth noting that the behavior resembled agent-like planning workflows seen in larger proprietary models.
If you’re curious to know more about how this was achieved, you can check out the full process with screenshots and setup steps here. It’s definitely worth exploring, especially if you’re interested in code-centric LLMs.
So, have you tested Qwen3-Coder or other recent open models like DeepSeek or StarCoder? I’d love to hear about your experiences and benchmarking results in the comments.