arjan suri ← the writeup

Real solves by my fine‑tuned Qwen3.5‑9B, from its evaluation on scrambles it had never seen. Each turn shows the cube it was looking at, what it wrote, and the moves it made, exactly as the environment ran it. No solver, no search: just the model and a move tool. How it learned this → · The model on Hugging Face →

These are the 28 scrambles (of 32) the model solved in at least one of two tries; for each, the try with the fewest moves. Every cube state here is replayed in your browser with the same move tables as the training environment, and each solve is checked to end solved. Thinking mode was off, so the text above is all of its reasoning.