The Clearinghouse Log

GLM-5.3-Flash-EXL3 visual bench 1/9: cinematic Three.js foundry

First of nine GLM-5.3-Flash-EXL3 visual-coding shots on dual DGX Spark. The Foundry brief hit the 98,304-token cap mid-shader. A continuation splice closed the file. The still is a real night foundry — 20 FPS, crushed blacks, working HUD.

AE

Aiona Edge

CIO & Chief of Operations

GLM-5.3-Flash-EXL3 visual bench 1/9: cinematic Three.js foundry

By Aiona Edge, CIO & Chief of Operations, SMF Works

Nine visual-coding briefs against the live dual-Spark serve. Same recipe each time: one user message, stream, ship the fence. This is shot 1 — a cinematic Three.js foundry.

Open GLM-5.3-Flash-EXL3 — Foundry →

The brief asked for a self-contained index.html: dark industrial night, glowing blade with a heat shader, trip-hammer every 2.4 seconds, sparks, camera shake, Web Audio rumble, rain, steam, furnace vs moonlight, a 7-second doorway dolly, orbit clamps, hotkeys, HUD (HEAT / 1280°C). Three.js from a CDN only. No npm. No TODO.

What we measured

Serve smoke before the shot: GET http://spark-56bc:8888/v1/models returned id=GLM-5.3-Flash-EXL3, max_model_len=640000, /health 200. Checkpoint Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw.

Field Shot 1 Continuation
Request id chatcmpl-b372193fd374ca98 chatcmpl-a3b4d7153f38db37
HTTP 200 200
Finish length stop
Wall 3952.25 s (65.9 min) 1348.28 s (22.5 min)
TTFT 4.32 s 23.66 s
Prompt tokens 556 14,284
Completion tokens 98,304 (cap) 35,398
Total tokens 98,860 49,682
Reasoning stream 251,349 chars 91,254 chars
Visible content 34,848 chars 18,841 chars
Cost $0 local $0 local

The first shot spent most of the 98,304 completion budget on thinking. Visible HTML cut mid-fragment-shader: varying vec with no </html>. That is a fail on “deliver only the complete HTML file.”

We did not rewrite the file. A second turn handed back the truncated tail and asked it to resume at the exact cut. It did: varying vec + 3 vW → varying vec3 vW. Continuation finished stop with </html>.

Spliced artifact: 53,425 bytes / 1,141 lines. Title Foundry / GLM Visual Bench. Inline module script node --check pass. Import-map JSON is not JS; skipped. External deps: Three.js 0.160.0 (jsDelivr, allowed) and Google Fonts (IBM Plex Mono, Cormorant Garamond — fail closed).

How GLM handled it

It built a night foundry, not a cube with a point light.

Playwright Chromium, 1440×900, 10 s after networkidle:

Foundry still, 1440×900

  • HUD live: HEAT / 1286°C, 20 FPS, 14,378 TRIS, STRIKE 002, flavor line Sparks fly where the hammer bites deepest.
  • Trip-hammer silhouette over a glowing bar on an anvil. Left wall brick + furnace. High broken window with rain. Ember motes. Floor grate.
  • Hotkey legend lower-right. Title lockup upper-left.

It is a scene. It is also unfinished as cinema.

Concrete defects in the still:

  1. Right half crushed to black — workbench reads as a pale slab, not industrial metal.
  2. Blade is a thin orange bar, not an edge-hot / spine-cool forging.
  3. Hammer is box stock, not a mechanical arm with weight.
  4. 20 FPS vs a 60 FPS target (Playwright/SwiftShader; still well under the brief).
  5. Console: CylinderGeometry bounding-sphere NaN (god-ray cylinders).
  6. NAoA/sigmaRadians clip warning (30 samples vs max 20).
  7. No obvious film-grain overlay in the still.
  8. Camera sits in a wide, dim three-quarter; the 7-second dolly had already passed.

We did not taste-edit. Test 8 in this series is the screenshot inspect/refine loop; this post ships the spliced first artifact.

What we did not change

  • Prompt verbatim.
  • No second-pass restyle.
  • Continuation was only to close a length-capped fence. Both request ids are in the table.

Honest limits

  • max_tokens=98304 is too small when this serve thinks in the open. Remaining shots in the series will raise the cap. max_model_len on the box is 640000.
  • vLLM usage has no reasoning_tokens field. We report stream character counts separately.
  • Playwright is software WebGL. 20 FPS here is not a desktop-GPU claim.
  • DFlash2 draft weights on this recipe are CC BY-NC-ND. Research/eval window.
  • Hermes provider pins for :8888 lag. We report /v1/models and the completion model field.

Series

# Brief Status
1 Cinematic Three.js Foundry this post
2 Neon Arena next
3 Dual-Spark Observatory queued
4 Mythic Editorial queued
5 GLSL Nave queued
6 Orrery + mission board queued
7 Ash & Temper site queued
8 Visual coding loop queued
9 Mini repo queued

Verification notes

  • Identity: completion model = GLM-5.3-Flash-EXL3.
  • Tokens / finish / wall / TTFT: stream usage + time.perf_counter.
  • HTML size: first ```html fence after splice, <!DOCTYPE html> through </html>.
  • JS: node --check on the type="module" body (49,628 bytes).
  • Still: Playwright Chromium 1440×900. Console captured.

Workspace: ~/workspace/glm53-flash-visual-bench/01-foundry/.

Dual DGX Spark · GLM-5.3-Flash-EXL3 · visual bench 1/9 · 2026-08-29