GLM-5.3-Flash EXL3
90,852 output tokens · 59 min · 26 tok/s · 57.9 KB · 0 console errors · fish visibly converge on food
Koi Pond: GLM vs MiMo on two Sparks
A busy koi pond with a waterfall, plants, and a feed button the fish respond to. Both models delivered first try. GLM's pond is clearly the better one; MiMo finished in half the time.
GLM-5.3-Flash EXL3 · TP2 · topMiMo-V2.6-Flash MXFP4 · 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 24 fps, so cadence is honest even where the scene would have run slower live.
MiMo-V2.6-Flash MXFP4
75,502 output tokens · 30 min · 42 tok/s · 20.9 KB · 0 console errors · pellets drop, koi barely visible under caustics
The setup, the caveats, the limits.
Prompt, verbatim: "Create a single HTML file using Three.js that renders a busy koi pond. Include a waterfall, plants, and multiple koi in different colors. Add a button that feeds the koi—the fish should swim toward the food and eat it. It should feel alive and be in continuous motion. Make it beautiful. Generate all geometry, materials, and effects in code. No external images, textures, or 3D models. Loading Three.js from a CDN is allowed."
Both ran with thinking on and no output cap on separate two-Spark pairs: GLM via EXL3 TP2, MiMo via vLLM TP2 using tonyd2wild's DGX Spark recipe (RoCE, fp8 KV, DFlash). One run each, zero console errors, no repairs. GLM: 90,852 tokens in 59 min (26 tok/s). MiMo: 75,502 tokens in 30 min (42 tok/s).
Recording method: 24 s fixed-step capture, Feed button clicked at 8 s and 16 s on both arms.
GLM: rock-rimmed pond, lily pads, cattails, visible orange/white/black koi that converge when fed. MiMo: square grey pond, boulder waterfall, fish lost under a noisy water texture. GLM wins on the artifact; MiMo wins the stopwatch.