Qwen 3.8 27B on a laptop nearly matches GPT-6 Astra at reading model weights

Simon Willison found a local Qwen 3.8 27B can name the game a small network plays from its raw weights, as GPT-6 Astra did in a new research post.

A model on your laptop can do one of the headline tricks from this week's weight-space programming research. In Sparks of Weight-Space Programming (Nov 2, 2026), the authors gave GPT-6 Astra the raw weights of small game-playing networks as text, with no tools, and it named the game each was trained on. Simon Willison reports on his weblog (Nov 3, 2026) that Qwen 3.8 27B, running locally, did the same for three of four games.

Qwen named Pac-Man after 46m 12s (72% confidence), Pong after 1h 4m 54s (75%) and Breakout after 52m 25s (88%). It got Freeway wrong. Kimi K3.1 Pro, run via OpenRouter, named Breakout in 9m 41s.

The prompts in the authors' GitHub logs are self-contained, so you can rerun them. Willison used his LLM command-line tool with the llm-lmstudio plugin. He had to raise LM Studio's default 8192-token context, which failed with errors, and lengthen the timeout:

# Willison raised the llm-lmstudio timeout with this variable
LMSTUDIO_TIMEOUT=900 \
llm -f https://raw.githubusercontent.com/RikiyaT/lm_understand-logs/refs/heads/main/exp1_which_game/runs/pacman/prompt.txt \
-m lmstudio/qwen/qwen3.8-27b

Keep the result in proportion. A daily.dev summary of one critique argues the game task is close to trivial, because the prompt labels each input channel and a three-line Python script computing L2 norms reveals the pattern. Willison himself expects a patient human with a notepad could get there.

The paper's bigger claim is that Astra lifted Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing weights. That number comes from the authors. Research & Development World notes that 93% of the base model's answers hit the 32,768-token limit. Critics Matthew Honnibal and Sean Goedecke argue the result came from the kind of Python analysis a machine-learning engineer would write. Honnibal says the announcement groups experiments with different tool access under one claim. Goedecke calls the fix an application of steering, an existing technique.

Sources