Wesche · DGX Spark LabOrigami Dragon Battle
Two models · two dual-Spark clusters · one uncapped prompt

One sheet of paper. Two living dragons.

GLM-5.3 Flash and Qwen3.8 Flash Next received the same demanding metamorphosis brief with thinking enabled and no output-token cap or wall-clock timeout. Drag inside either scene to orbit, scroll to zoom, use the authored camera presets, or compare both live.

Read the exact generation prompt

Create a stunning, cinematic, interactive 3D animation as one complete HTML file: ORIGAMI DRAGON METAMORPHOSIS. Begin with a single square sheet of metallic paper floating above a carved stone pedestal in a moonlit Japanese temple courtyard. The paper must visibly fold through multiple coherent stages—creases, angular intermediate forms, articulated head, wings, legs and tail—until it becomes an elegant, highly detailed origami dragon. Do not hide the transformation with a cut, flash, fade or camera obstruction: the folding process itself must remain clearly visible and spatially continuous. After forming, the dragon must awaken, unfold and articulate its wings, launch from the pedestal, fly a complete cinematic path around the courtyard, bank and turn naturally, move its segmented tail, and breathe a luminous stream of fire made from thousands of dynamic particles that illuminate the dragon and environment. Add drifting embers, paper fibers, mist, moonlight, moving shadows, reflective metallic folds, torii gates, stone lanterns and wind-blown foliage. The dragon must return to the pedestal and visibly unfold back into the original square sheet for a seamless loop. Technical and interaction requirements: - True 3D perspective rendered with WebGL; Three.js from a public CDN is allowed, but all scene code, shaders, styles and UI must be contained in this one HTML file. - Use procedural geometry only: no downloaded 3D models, images, textures, video or other external art assets. - Use at least 20,000 animated particles across fire, embers, atmospheric dust and transformation accents. - Smooth animation targeting 60 fps with adaptive quality if necessary. - Cinematic bloom, depth, fog, shadows and color grading. - Slow automatic camera movement plus mouse/touch drag to orbit and wheel/pinch to zoom. - Include camera presets, including a button named HERO VIEW, and a subtle HUD showing FPS, particle count, camera azimuth and elevation. - Make the composition beautiful, mysterious and production-quality on desktop and mobile. Output only the complete HTML file in a single fenced html block, with no omitted sections or placeholder code.

2 × TP2dual-DGX-Spark deployments
20,000+particles requested
Uncappedgeneration budget
Interactivedrag · zoom · presets
Open GLM full-screen ↗
GLM-5.3 Flash · Origami DragonOriginal output · rendered untouched
GLM-5.3 FlashOriginal render
Qwen3.8 Flash NextOne display-only repair

Interactive: drag/touch the scene, wheel or pinch to zoom, and use each model’s own HUD and camera presets.

GLM-5.3 Flash · DFlash2 · TP2 · xhigh

Moonlit metallic dragon

Reasoning 66,961 tokGenerated 81,179 tokWall 53m 56s

The model’s original HTML executes without repair. Reasoning tokens were reconstructed with the native GLM tokenizer over the preserved reasoning stream.

Original model output rendered untouched.

Open standalone GLM scene →

Qwen3.8 Flash Next · NEXTN · TP2 · max

Temple awakening and flight

Reasoning 117,329 tokGenerated 153,000 tokWall 96m 43s

Qwen’s endpoint reported its reasoning count directly. The generation completed normally, but its HTML contained one malformed, unused JavaScript light declaration that prevented parsing.

Disclosure: the live copy removes that single unused declaration. No geometry, animation, materials, timing, lighting design, or scene concept was rewritten.

Open repaired Qwen scene →

Benchmark integrity

The immutable first-run HTML files are preserved separately. The web gallery serves a clean original for GLM and a clearly disclosed display copy for Qwen.

Equivalent conditions

Both requests ran concurrently on TP2 model pairs with thinking enabled, the same prompt, no request token cap, and no wall-clock timeout.

Matched-camera comparison

Both scenes are stacked at AZ 35.5° / elevation 24.7°. The formed-dragon sequence is slowed equally for visibility; capture exposure is normalized and disclosed in-frame.