GLM-5.3 FlashTwo-line display repair
Flash TP2 vs Full 743B TP4 · same uncapped prompt
One solar system. Two very different generation paths.
GLM-5.3 Flash NVFP4 on two DGX Sparks and Full GLM-5.3 743B Int4/Int8Mix on four DGX Sparks received the exact same 422-token brief. Thinking was enabled, temperature was 0.7, top-p was 0.95, and neither request had an output-token cap or wall-clock timeout.
Read the exact generation prompt
Create an exceptionally polished interactive solar system experience as one complete self-contained HTML document. Return only the HTML document, beginning with <!DOCTYPE html> and ending with </html>. The experience must run fully offline in a modern browser with no external libraries, images, fonts, APIs, or network requests. Use only embedded HTML, CSS, JavaScript, Canvas 2D, WebGL, or inline SVG. Design a cinematic full-screen deep-space scene containing the Sun and all eight planets. Make each planet immediately recognizable through color, texture, rings, atmosphere, lighting, and scale while keeping the system visually usable rather than claiming literal astronomical scale. Include elliptical orbital paths, visibly different relative orbital speeds, Earth's Moon, several major moons around Jupiter, Saturn's rings, a dense asteroid belt, a layered starfield, and an occasional comet with a glowing tail. The solar system must continuously animate and feel spatial rather than like a flat diagram. Add smooth perspective depth, orbital inclination, planet rotation, glow, shadows, subtle bloom-like effects, and polished transitions while maintaining good frame rate. Required interactions: - drag or swipe to rotate the viewing angle; - mouse wheel or pinch-style controls to zoom; - click or keyboard-select any major body to smoothly focus the camera on it; - show a well-designed factual information panel for the selected body; - play/pause control; - simulation-speed control with a wide useful range; - toggles for labels, orbital paths, and motion trails; - an overview/reset control; - a clear way to leave focused-follow mode and return to the full system. Make the interface responsive on desktop and mobile, visually coherent, accessible by keyboard, and understandable without instructions. Use elegant typography, restrained glass-like panels, high contrast, and no visual clutter. Include a compact title and a small unobtrusive note that sizes and distances are artistically adjusted for visibility. Prioritize working interaction and animation over decorative complexity. Avoid placeholder elements, fake controls, static screenshots, broken features, and explanatory prose outside the HTML.
2 vs 4DGX Sparks · TP2 vs TP4
29,945 vs 82,364thinking tokens*
31:09 vs 1:21:57wall time
Uncappedboth generations
GLM-5.3 Flash · Interactive Solar SystemDisplay copy · disclosed two-line camera/Sun repair
Full GLM-5.3 743BUnmodified output
Interactive: drag or swipe, scroll to zoom, select planets, change simulation speed, toggle labels/orbits/trails, and return to overview.
GLM-5.3 Flash · NVFP4 · DFlash2 K7 · TP2
Fast plan, broken projection
Thinking* 29,945 tokGenerated 46,418 tokTotal 46,840 tok24.94 tok/s31m 09s
The request completed normally with finish_reason: stop, but the native visual placed the main bodies effectively off-camera and omitted the Sun's projected radius.
Disclosure: the live copy adds camera distance to finite-body projection depth and assigns the Sun's projected radius. The untouched original remains available and retains its
fail_major visual verdict.Full GLM-5.3 743B · Int4/Int8Mix · MTP K5 · TP4
Long plan, clean first render
Thinking* 82,364 tokGenerated 105,787 tokTotal 106,209 tok21.52 tok/s1h 21m 57s
Full GLM completed normally and its untouched 66 KB HTML passed document, JavaScript, offline, browser-runtime, control, and motion validation with zero browser errors.
Original model output rendered untouched. No display repair was needed.
What the numbers mean
*The APIs reported aggregate completion usage but did not split reasoning tokens. Thinking counts are exact tokenizer counts over each preserved reasoning stream using the GLM tokenizer. They reconcile with native completion totals within two or three special tokens.
Deployment comparison
This is Flash NVFP4 TP2 on two Sparks versus Full 743B Int4/Int8Mix TP4 on four Sparks. It measures these deployments under one creative request, not universal model superiority.
Stacked comparison with full metrics
Flash is shown on top with its repair badge; Full is shown below unmodified. Quant, topology, thinking/generated/total tokens, throughput, wall time, and finish status remain visible in-frame.
Flash standalone capture
Disclosed two-line display repair.
Full 743B standalone capture
Unmodified model output.
Public evidence
Machine-specific networking and private paths are excluded. The generation contract, exact prompt, hashes, metrics, repair details, native failure verdict, and Full browser validation are published here.