Qwen3.6 27B local vs Opus 4.8, voxel engine in raw C with zero frameworks

A comparison experiment pitted Claude Code on Opus 4.8 against a locally running Qwen3.6 27B model to build a voxel world engine in plain C without any external frameworks or libraries.

• The local Qwen3.6 27B was executed via vLLM using NVFP4 quantization and a 256k context window, achieving approximately 130 TPS on an RTX 6000 Blackwell 96GB.
• Opus 4.8 demonstrated superior understanding of voxel physics, maintaining terrain integrity, chunk alignment, and functional collision detection.
• Qwen3.6 27B successfully compiled and rendered the code but suffered from memory management issues that caused visual tearing, despite handling raw pointers and manual allocation.

The results highlight a significant improvement in local models' ability to handle complex C programming tasks, suggesting that the performance floor for open-weight models is rising rapidly even if frontier models remain ahead.