AICodeKing tests Step 5 Preview through OpenCode on eight coding and reasoning tasks, inspecting the resulting projects rather than relying only on a benchmark score. An elevator simulation, a lens case and a folding table work in part but expose interaction and geometry bugs. An SVG panda and a playable archery game provide stronger results.
AICodeKing also demonstrates a local fine-tuning workflow using a small Gemma model, LoRA adapters and a web interface. The workflow needs a continuation after a cache-configuration issue, while factual errors in the generated training data remain a quality limitation. A wristwatch rendering performs less well even after regeneration.
AICodeKing awards the preview 67 out of 80 points and compares that result with earlier GLM and MiMO tests. Those comparisons belong to this creator's task suite, not a controlled universal model ranking. The episode illustrates why working demonstrations, bug inspection and training-data review matter alongside aggregate scores.
Watch on YouTube
