Qwen 3.8 27B runs on a laptop and reads raw model weights to guess the game

Key points

• Simon Willison ran Qwen 3.8 27B locally in LM Studio on a laptop, testing the paper's no-tools weight-reading task.
• On the Pac-Man weights, the model took 46m 12s and 43,506 reasoning tokens, guessing an 8x8 maze game.
• It reportedly identified Pong after 1h 4m 54s, while it was still working on the Freeway prompt at writing time.
• The model is about 17GB and Willison says the paper's approach needs no tools, only reasoning over numbers.

Why it matters for marketers: Content teams testing local models should expect long reasoning times and context settings to be tuned before results are usable.