DeepSeek V4.1 Flash EXL3
98,931 output tokens (65,413 reasoning) · 51 min · 28 tool calls · 35.8 KB · hyperspace, pause, attract mode
Asteroids: Two Local Models, One Memorized Game.
Same short prompt, ultra reasoning, two 2-Spark pairs. The results are ~90% the same game — which says more about the prompt than the models. A fully-specified 1979 classic tests recall, not design.
DeepSeek V4.1 Flash EXL3 · TP2 · topGLM-5.3 Flash EXL3 · TP2 · bottom
Panel order matches the labels above. These are headless-browser recordings of the generated code driven by a scripted, identical input timeline — not recordings of model inference. Fixed-step virtual clock at 30 fps, so cadence is honest even where the scene would have run slower live.
GLM-5.3 Flash EXL3
97,648 output tokens · 70 min · 55 tool calls · 25.8 KB · saved high score, pause-on-blur, wave speed scaling
The setup, the caveats, the limits.
Prompt, verbatim: "Build a playable Asteroids game as a single self-contained HTML file (inline CSS/JS, no external libraries, assets or network). Capture the 1979 Atari feel: rotate/thrust, shoot and split asteroids, screen wrap, lives, score, waves, game over and restart. Add any polish you think makes it feel great. Output only the HTML file in one code block."
Both ran under Hermes with /reasoning ultra, uncapped output, from the same fresh session state. DeepSeek finished in 59 minutes, GLM in 78. GLM spent roughly a third of its run on a headless-Chrome smoke test with mutator hooks before writing the clean file.
Recording method: same simulated keyboard schedule on both. Each game seeds its own asteroid field, so ship paths differ even with identical inputs. Both scored 1000+ and lost a life inside 40 seconds.
What this proves: not much about the models. Asteroids is in every training set. The follow-up lighthouse study used an underspecified Three.js brief for exactly this reason.