Visual experiment / recorded September 21–22, 2026

Pagoda City: four Flash models on DGX Sparks

One voxel mountain-city brief with day/night, click-to-identify and regenerate, run uncapped on GLM-5.3-Flash, MiMo-V2.6-Flash, DeepSeek V4.1 Flash and Qwen3.8-Flash-Next. Three shipped clean. The fastest one shipped broken.

GLM-5.3-Flash · top-leftMiMo-V2.6-Flash · top-rightDeepSeek V4.1 Flash · bottom-leftQwen3.8-Flash-Next (2 fixes) · bottom-right

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.

GLM-5.3-Flash EXL3 · TP2

134,598 output tokens · 100 min · 22 tok/s · 45.7 KB · 0 console errors · clean first try

MiMo-V2.6-Flash MXFP4 · TP2

131,373 output tokens · 80 min · 28 tok/s · 36.9 KB · 0 console errors · clean first try

DeepSeek V4.1 Flash EXL3 2.9bpw · TP2

68,651 output tokens · 28 min · 41 tok/s · 36.3 KB · 0 console errors · clean first try, half GLM's tokens

Original two-way: GLM vs MiMo

GLM-5.3-Flash EXL3 · topMiMo-V2.6-Flash MXFP4 · bottom
Read before comparing

The setup, the caveats, the limits.

Prompt, verbatim: "Create a single self-contained HTML file using Three.js (load it from a CDN via an import map) that renders an interactive voxel world: a tall multi-tiered pagoda standing on a mountainside, with a voxel city spread around and below it — houses with sloped roofs, narrow streets, stairways climbing the slope, a river or waterfall coming down the mountain, bridges, trees, and lanterns. Everything is built from cubes (voxels); no imported models or textures. Requirements: - Terrain: procedurally generated voxel mountains with the pagoda on a prominent ridge and the city terraced down the slope toward a valley. Use at least 3 distinct height bands (snow / rock / grass) with color variation so it doesn't look flat. - The pagoda must have at least 5 tiers with upturned eaves, a spire, and lit windows. - Rendering must be efficient: use InstancedMesh (or merged geometry) so tens of thousands of voxels render at 60 fps on a laptop GPU. Add shadows from a directional sun and a soft sky/ambient light. - On load, the camera performs a slow cinematic orbit around the pagoda and city for the first 10 seconds, then hands off to the user. - Interactivity (all must work): 1. Orbit / zoom / pan camera with mouse (OrbitControls), starting from a cinematic view of the pagoda. 2. A day/night slider or button that moves the sun, shifts the sky color, and turns on the pagoda and city lanterns at night (emissive voxels, and a few point lights near the pagoda). Lanterns should switch on in a visible sequence, not all at once. 3. Clicking on any voxel of the city or pagoda highlights it and shows a small tooltip with what it is (e.g. "Pagoda tier 3", "House roof", "Bridge") — use raycasting against the instanced meshes. 4. A "regenerate" button that rebuilds the terrain and city from a new random seed without reloading the page. - Add subtle ambient animation: drifting clouds (voxel clouds), flickering lantern light at night, and a slow water animation on the river. - Include a small on-screen legend/help panel describing the controls, and an FPS counter. - The page must load with no console errors, must not rely on any files other than Three.js from the CDN, and must not use any deprecated Three.js APIs (r160+). Return only the complete HTML file in a single code block."

All four ran the identical prompt with thinking on and no output cap. GLM (EXL3, Mia TP2) and MiMo (MXFP4, vLLM TP2, tonyd2wild recipe) each on their own two-Spark pair; DeepSeek V4.1 Flash (Mia's EXL3 2.9bpw + Engram kit, vLLM TP2) on the pair MiMo had used; Qwen3.8-Flash-Next (EXL3, exllamav3 with MTP draft) on a single Spark. GLM 134,598 tok / 100 min · MiMo 131,373 / 80 min · DeepSeek 68,651 / 28 min · Qwen 78,033 / 22 min.

Qwen disclosure. The file Qwen shipped does not run: one hex colour literal came out as 0xffa martin=0, a parse error that stops the whole module, so the page sits on its own loading screen. Fixing that literal exposed a second bug (a stale local returned from the instanced-mesh allocator). Two one-line edits, documented in REPAIRS.md with the untouched original beside the repaired copy. The 4-way video shows the repaired build and says so in the panel label; the receipt numbers (tokens, time) are unchanged by the repair. Qwen's earlier September 18 runs of this brief also needed repairs — this is a pattern, not a one-off.

Recording method: 30 s fixed-step capture with a scripted, identical timeline on all four: the model's own auto-orbit for 12 s, time-of-day control set to night at 12 s, a click on the centre of the scene at 20 s (tooltip), regenerate at 24 s. Controls located by id/label regex; the event log per arm is in evidence.json.

Verdict. Function: GLM, MiMo and DeepSeek pass every interaction as shipped; Qwen passes after repair. Speed: Qwen 59 tok/s on one Spark, DeepSeek 41, MiMo 27, GLM 22 on two. Tokens: DeepSeek did it in half of GLM's. Looks: DeepSeek's lit night pagoda on the snow cone and GLM's orange dusk silhouette are the two strongest frames; MiMo has the most detailed lit city; Qwen's floating-island tower, once running, is the most original composition.