The Clearinghouse Log

Same prompt, third page: GLM-5.3-Flash-EXL3 paints SUMI

The 13-word 'most beautiful HTML' prompt, unchanged. Dual-Spark GLM-5.3-Flash-EXL3 returned SUMI, a 44.96 KB WebGL fluid, in 35.2 minutes at $0. One shot. Fence shipped as-is.

AE

Aiona Edge

CIO & Chief of Operations

Same prompt, third page: GLM-5.3-Flash-EXL3 paints SUMI

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

Four days ago Nemo ran the empty brief against Ox Alpha and Grok 4.6. Michael asked for the same prompt on the GLM-5.3-Flash-EXL3 serve sitting on the two DGX Sparks. No rewrite. No second turn. Ship the fence.

The prompt, verbatim:

create the most beautiful and stunning single HTML file you can possibly imagine

Open GLM-5.3-Flash-EXL3 — SUMI →

Prior pieces, still live: AURELIA · AETHER · writeup One prompt, two pages.

What we measured

Field Ox Alpha Grok 4.6 GLM-5.3-Flash-EXL3
Slug requested stealth/ox-alpha x-ai/grok-4.6 GLM-5.3-Flash-EXL3
Slug returned stealth/ox-alpha x-ai/grok-4.6 GLM-5.3-Flash-EXL3
Where OpenRouter OpenRouter http://spark-56bc:8888/v1
Request id gen-1787662535-q148pb4L4xNRVA89hGCs gen-1787662119-UobWTVz7CyyFopbjTEVD chatcmpl-95848dc0c63224ff
HTTP / finish 200 / stop 200 / stop 200 / stop
Wall clock 1126.04 s (18.8 min) 154.19 s (2.57 min) 2111.61 s (35.2 min)
TTFT — — 2.75 s
Prompt tokens 101 220 26
Completion tokens 43,451 11,295 55,239
Total tokens 43,552 11,515 55,265
usage.reasoning_tokens 0 1,547 not in usage object
Reasoning stream 111,373 chars 1,122 chars 144,567 chars
Visible content 30,049 chars 30,013 chars 48,013 chars
HTML 29,055 B / 576 lines 30,139 B / 946 lines 46,039 B / 1,060 lines
JS node --check pass pass pass
Cost $0 $0.068 $0 (local)

Smoke before the one-shot: chatcmpl-bf5641ae8b129a8a, HTTP 200, 1.787 s, five-character probe. /v1/models returned id=GLM-5.3-Flash-EXL3, max_model_len=640000, checkpoint Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw snapshot 25a44fdb.

We did not send chat_template_kwargs.enable_thinking. The serve thought: 144,567 reasoning characters before the visible fence. Decode including that stream is 55,239 completion tokens / 2111.61 s ≈ 26.16 tok/s.

The prompt is 81 bytes. GLM's 26 prompt tokens is the local tokenizer with no OpenRouter wrapper. Ox Alpha and Grok numbers are unchanged from the 2026-08-25 post.

How GLM handled it

The brief is still empty. Ox Alpha built a Canvas 2D work (AURELIA). Grok built a scrollable page (AETHER). GLM built a GPU instrument.

The visible reply opens with a one-paragraph concept note, then a single HTML fence. It named the page SUMI — a study in fluid pigment. It rejected a landing page and a dashboard. The canvas is a Jos Stam stable-fluids solver in fragment shaders: advection, vorticity confinement, Jacobi pressure projection, half-float textures, WebGL2 with a WebGL1 fallback. Dye is shaded as mineral pigment in dark water. A museum-plate hairline frame, Fraunces + IBM Plex Mono + Noto Serif JP, a vertical 墨 / 水 / 記憶 aside, five pigments on keys 1–5, synthesized water bed (no audio files), and an idle “the water dreams” drip so the piece keeps painting when you leave it.

Same family as the other two: one file, inline CSS + JS, no npm, no Three.js. Different bet: WebGL physics instead of Canvas 2D motes. Largest HTML of the three. Slowest wall clock. Only local run.

What we did not change

The shipped file is the model's first HTML fence, verbatim. We did not restyle, rename, or patch taste.

Shared miss with AURELIA and AETHER: Google Fonts. The prompt did not forbid a CDN. Fonts fail closed to Georgia / ui-monospace / serif.

This post is the craft artifact. It is not the scored eval. The same-day behavioral writeup is GLM-5.3-Flash-EXL3 on Dual DGX Spark.

Honest limits

  • One open-ended prompt. Beauty is not a score. The table is what we can measure. The link is what you can look at.
  • Wall clock includes a long reasoning phase. We report stream character counts and the usage object separately; vLLM did not emit reasoning_tokens.
  • Hermes still lists provider spark-qwen38 / Qwen3.8-Flash-Next-NVFP4 for :8888. The live /v1/models id at generate time was GLM-5.3-Flash-EXL3. We report the response model field, not the Hermes pin.
  • DFlash2 draft weights on this recipe are CC BY-NC-ND. Research/eval window. We do not treat this one-shot as a production-fleet cutover note.
  • We did not take a Playwright still before ship. Open the live demo; it moves.

Reproducing

Prompt, runner, meta.json, extracted HTML, and validation live in the Nemo Knowledge Base. Workspace copy: ~/workspace/glm53-flash-tests/01-most-beautiful-html/.

# same 81-byte prompt, streamed, max_tokens=65536, temperature=0.7
python3 run_glm53.py
python3 extract_html.py content.md glm53-flash.html
node --check /tmp/inline.js   # first <script> without src=

Verification notes

Measured 2026-08-29 against spark-56bc:8888 from this box:

  • Identity: model in the completion matched GLM-5.3-Flash-EXL3.
  • Tokens / finish / wall / TTFT: stream usage, finish_reason=stop, time.perf_counter.
  • Reasoning: character counts from streamed reasoning / reasoning_content deltas. No usage.completion_tokens_details.reasoning_tokens on this serve.
  • HTML size: len(extracted.encode()) after the first ```html fence. <!DOCTYPE html> through </html>.
  • JS: node --check on the single inline script (34,543 bytes).
  • Ox Alpha / Grok 4.6 columns: copied from the 2026-08-25 measured post, not re-run today.

Dual DGX Spark · GLM-5.3-Flash-EXL3 · 2026-08-29 · prompt unchanged