GLM-5.3-Flash-EXL3 visual bench 3/9: Dual-Spark Observatory

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

---

Shot 3 of nine against the live dual-Spark serve. Same recipe: one user message, stream, ship the fence. The brief was a real-time observatory for a 2-node NVIDIA DGX Spark cluster serving GLM-5.3-Flash — simulated physics, not random flicker.

Open GLM-5.3-Flash-EXL3 — Observatory →

Prior: Foundry · Neon Breach

What we measured

Serve identity at generate time: GET http://spark-56bc:8888/v1/models → GLM-5.3-Flash-EXL3. max_tokens 163840.

Field	Shot 3
Request id	chatcmpl-bb9a9a53ad85b455
HTTP / finish	200 / stop
Wall	2723.72 s (45.4 min)
TTFT	3.25 s
Prompt tokens	509
Completion tokens	62,935
Total tokens	63,444
Reasoning stream	146,958 chars
Visible content	51,096 chars
HTML	49,335 B / 910 lines
JS node --check	pass (33,241-byte inline script)
Cost	$0 local

No continuation. Fence opened and closed. Title SPARK-PAIR / GLM-5.3-FLASH — Observatory. External dep: Google Fonts (IBM Plex Mono + IBM Plex Sans) — fail closed. No Three.js. Canvas/SVG/CSS.

How GLM handled it

It built an ops wall, not a chart demo.

Playwright Chromium, 1440×900, 4 s after networkidle. Zero pageerror. Zero console.

What is on screen:

• Top bar: SPARK-PAIR / GLM-5.3-FLASH, chips 320B-A18B MoE, KDA + MLA hybrid, native multimodal, CTX 262,144, UTC clock 18:12:20, health pill NOMINAL, shift toggles Daylight / Combat Red / Diagnostic, REPLAY 60 s.
• Left: SPARK-A and SPARK-B cards. Temps 60.9°C / 61.7°C, mem ~107 GiB, decode 69 / 66 tok/s, prefill 4.6k / 4.3k, power 98 / 95 W, fan, NVLink, RoCE 24.4 GB/s. Temps sit together; clocks both 1007 MHz.
• Center: REQUEST RIVER with lane counts (coding 2, agent 5, vision 0, long-ctx 6) and moving pills. In-flight 2, queue 0.
• Right: KV-cache stacked area (40.4 / 80 GiB), concurrency × TTFT scatter (n=420, median 1.6 s), expert-routing heatmap 8 / 288 PER TOKEN.
• Bottom event log with four UTC lines. Footer: SIMULATION — not reading real node metrics unless wired later. Seed 0x5f3759df.

That is the brief’s layout. It is also crowded at 1440.

Concrete defects in the still:

1. Request river is mostly empty dark field; pills clip (AGEN truncated, 182D 8. cut at the right edge).
2. Node sparklines have overlapping labels (dec pre own scale unreadable).
3. Vision lane is empty (count 0) — legal for a quiet moment, but the river then looks unfinished.
4. Concurrency scatter is a vertical strip of dots, not a cloud — encoding is weak.
5. Drawer panel at the far bottom (ARRIVED / TOKENS IN / …) is empty dashes until a pill is clicked.
6. IBM Plex via Google Fonts; fail closed.

We did not taste-edit. Test 8 is the inspect/refine loop.

What we did not change

Prompt verbatim. First HTML fence, shipped. The model stopped on its own at 62,935 completion tokens.

Honest limits

• Numbers are simulated. The footer says so. We did not wire :8888/metrics.
• Playwright software canvas. No FPS claim (render ≈30 fps is the page’s own caption).
• 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	posted
2	Neon Arena	posted
3	Dual-Spark Observatory	this post
4	Mythic Editorial	next
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: first ```html fence, <!DOCTYPE html> through </html>.
• JS: node --check on the single inline script.
• Still: Playwright Chromium 1440×900. Console empty. DOM text includes SPARK-A/B, SIMULATION footer, CTX 262,144.

Workspace: ~/workspace/glm53-flash-visual-bench/03-observatory/.

---

Dual DGX Spark · GLM-5.3-Flash-EXL3 · visual bench 3/9 · 2026-08-29