Qwen3.6 one-shots flashy Tetris clone

This is not a benchmark, but it is the kind of artifact developers actually care about: can the model turn a loose prompt into a working, polished web app without a repair loop?

• –Qwen3.6-35B-A3B is a 35B sparse MoE model with roughly 3B active parameters, making local coding demos unusually practical for its capability class
• –The generated output took 9,113 tokens and about 11 minutes, so quality came with real latency even at 13.62 tokens/sec
• –Unsloth's GGUF path matters because it lowers the hardware barrier for local experimentation, especially with llama.cpp-style serving
• –A flashy Tetris clone is toy-scale, but frontend coherence, game state, animation, and particle effects are a useful stress test for one-shot code generation