Ciru Inference Labllm.ciru.ai / research

Gemma4 26B QAT MTP 32-slot artifact run

A single local Strix Halo serving run launched 32 concurrent workers against Gemma4 26B QAT with draft-MTP enabled. Each worker generated a complete single-file HTML page from a different prompt. This page preserves the generated artifacts, screenshots, prompts, and measured run data.

Workers done32/32

Generated tokens71,637

Active weighted TG/s241.9

MTP accepted87.6%

Avg TTFP3.89s

Output chars252,191

Run configuration

Artifact setfirst complete 32-worker artifact run

Profilegemma4-26b-p32-mtp

Modelgemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf

Draft modelmtp-gemma-4-26B-A4B-it.gguf

Backendllama.cpp Vulkan / RADV Strix Halo

Slots32 parallel slots

Context262,144 total / 8,192 per slot

Generation cap7,680 tokens per worker

MTP settingsdraft-mtp, n-max 2, p-min 0.0

KV cachetarget f16/f16, draft f16/f16

Batchb4096 / ub1024

Threads16 / 16 batch

Samplingtemp 0.6, top-p 0.95, top-k 20

Artifact audit

All 32 workers finished with status done. The saved HTML artifacts contain HTML wrappers, CSS, and JavaScript. The post-run scan found no saved channel markers, no obvious repeated-token loop pattern, and no missing artifact files. Active weighted TG/s sums each worker's generated-token rate over its own active runtime.

Run source

This page uses the first complete 32-worker artifact run. Worker 07 is flagged only because it is a live WebGL canvas artifact.

Generated pages

Worker 01

Create a complete single-file HTML landing page for a futuristic local AI inference lab

Worker 02

Create a complete single-file HTML canvas game where the player catches falling stars

Worker 07 WebGL page

Create a complete single-file HTML Three

Worker 08

Create a complete single-file HTML kinetic typography animation

Worker 09

Create a complete single-file HTML dashboard for monitoring 32 AI workers

Worker 10

Create a complete single-file HTML mini game called Pulse Runner

Worker 11

Create a complete single-file HTML Three

Worker 12

Create a complete single-file HTML music-visualizer style animation without audio input

Worker 13

Create a complete single-file HTML product page for a local LLM appliance

Worker 14

Create a complete single-file HTML game called Target Pop

Worker 15

Create a complete single-file HTML Three

Worker 16

Create a complete single-file HTML animated explainer page about prompt processing vs token generation

Worker 17

Create a complete single-file HTML control panel for a spaceship

Worker 18

Create a complete single-file HTML game called Memory Grid

Worker 19

Create a complete single-file HTML Three

Worker 20

Create a complete single-file HTML looping title-card video for a model demo

Worker 21

Create a complete single-file HTML SaaS analytics homepage

Worker 22

Create a complete single-file HTML game called Keyboard Pilot

Worker 23

Create a complete single-file HTML Three

Worker 24

Create a complete single-file HTML interactive timeline page

Worker 25

Create a complete single-file HTML finance-style market dashboard

Worker 26

Create a complete single-file HTML game called Color Reflex

Worker 27

Create a complete single-file HTML Three

Worker 28

Create a complete single-file HTML animated comic panel sequence

Worker 29

Create a complete single-file HTML documentation page for a local benchmark run

Worker 30

Create a complete single-file HTML game called Circuit Clicker

Worker 31

Create a complete single-file HTML Three

Worker 32

Create a complete single-file HTML cinematic loading animation for a 32-slot model demo

Prompts and per-worker measurements

Each thumbnail above opens the generated single-file page. The full prompt and per-worker run data are preserved here so the gallery can stay focused on the 32 visual outputs.

Worker 012,609 generated tokens

Create a complete single-file HTML landing page for a futuristic local AI inference lab. Output code only starting with <!doctype html>. Include CSS, responsive sections, cards, and a tiny JavaScript status ticker. Finish the file with </html>.

Generated tokens: 2,609
Output: 8,805 chars
Total time: 357.455 s
TTFP: 3.858 s
Generated speed: 7.30 tok/s
Draft accepted: 1,657 / 1,904 (87.0%)

Worker 022,997 generated tokens

Create a complete single-file HTML arcade reaction game called Neon Dash. Output code only starting with <!doctype html>. Include CSS, score, timer, start button, keyboard controls, and a JavaScript game loop. Finish the file with </html>.

Generated tokens: 2,997
Output: 10,602 chars
Total time: 375.834 s
TTFP: 3.858 s
Generated speed: 7.97 tok/s
Draft accepted: 1,941 / 2,110 (92.0%)

Worker 031,610 generated tokens

Create a complete single-file HTML page using Three.js from a CDN. Output code only starting with <!doctype html>. Make a rotating neon cube field with camera animation and visible controls text. Finish the file with </html>.

Generated tokens: 1,610
Output: 5,440 chars
Total time: 209.433 s
TTFP: 3.850 s
Generated speed: 7.69 tok/s
Draft accepted: 1,020 / 1,182 (86.3%)

Worker 041,966 generated tokens

Create a complete single-file HTML animated video-style page. Output code only starting with <!doctype html>. Use CSS keyframes and JavaScript to animate a countdown, moving shapes, captions, and a looping progress bar. Finish the file with </html>.

Generated tokens: 1,966
Output: 7,370 chars
Total time: 274.716 s
TTFP: 4.017 s
Generated speed: 7.16 tok/s
Draft accepted: 1,238 / 1,454 (85.1%)

Worker 052,251 generated tokens

Create a complete single-file HTML portfolio page for a Strix Halo workstation. Output code only starting with <!doctype html>. Include responsive CSS, benchmark cards, feature rows, and simple filtering JavaScript. Finish the file with </html>.

Generated tokens: 2,251
Output: 8,259 chars
Total time: 315.913 s
TTFP: 3.844 s
Generated speed: 7.13 tok/s
Draft accepted: 1,433 / 1,636 (87.6%)

Worker 061,711 generated tokens

Create a complete single-file HTML canvas game where the player catches falling stars. Output code only starting with <!doctype html>. Include score, lives, timer, keyboard movement, and restart logic. Finish the file with </html>.

Generated tokens: 1,711
Output: 6,308 chars
Total time: 217.407 s
TTFP: 4.020 s
Generated speed: 7.87 tok/s
Draft accepted: 1,102 / 1,218 (90.5%)

Worker 071,445 generated tokens

Create a complete single-file HTML Three.js scene from a CDN. Output code only starting with <!doctype html>. Build a rotating torus knot, particle stars, mouse parallax, and a small FPS-style overlay. Finish the file with </html>.

Generated tokens: 1,445
Output: 4,740 chars
Total time: 179.760 s
TTFP: 3.867 s
Generated speed: 8.04 tok/s
Draft accepted: 928 / 1,032 (89.9%)

Worker 081,642 generated tokens

Create a complete single-file HTML kinetic typography animation. Output code only starting with <!doctype html>. Animate words about local inference, tokens, slots, and throughput using CSS and JavaScript. Finish the file with </html>.

Generated tokens: 1,642
Output: 6,119 chars
Total time: 216.993 s
TTFP: 3.858 s
Generated speed: 7.57 tok/s
Draft accepted: 1,034 / 1,216 (85.0%)

Worker 092,550 generated tokens

Create a complete single-file HTML dashboard for monitoring 32 AI workers. Output code only starting with <!doctype html>. Include CSS grid cards, live-looking counters, sparklines drawn on canvas, and update logic. Finish the file with </html>.

Generated tokens: 2,550
Output: 8,594 chars
Total time: 346.064 s
TTFP: 3.849 s
Generated speed: 7.37 tok/s
Draft accepted: 1,641 / 1,816 (90.4%)

Worker 102,327 generated tokens

Create a complete single-file HTML mini game called Pulse Runner. Output code only starting with <!doctype html>. Include a player square, obstacles, collision detection, score, speed ramping, and restart button. Finish the file with </html>.

Generated tokens: 2,327
Output: 7,882 chars
Total time: 323.375 s
TTFP: 3.851 s
Generated speed: 7.20 tok/s
Draft accepted: 1,490 / 1,674 (89.0%)

Worker 112,024 generated tokens

Create a complete single-file HTML Three.js solar system toy from a CDN. Output code only starting with <!doctype html>. Include orbiting planets, labels, camera controls via mouse movement, and animation loop. Finish the file with </html>.

Generated tokens: 2,024
Output: 7,287 chars
Total time: 293.867 s
TTFP: 4.039 s
Generated speed: 6.89 tok/s
Draft accepted: 1,256 / 1,534 (81.9%)

Worker 122,037 generated tokens

Create a complete single-file HTML music-visualizer style animation without audio input. Output code only starting with <!doctype html>. Use canvas bars, waveform motion, CSS controls, and random beat pulses. Finish the file with </html>.

Generated tokens: 2,037
Output: 7,020 chars
Total time: 288.710 s
TTFP: 3.857 s
Generated speed: 7.06 tok/s
Draft accepted: 1,280 / 1,512 (84.7%)

Worker 133,311 generated tokens

Create a complete single-file HTML product page for a local LLM appliance. Output code only starting with <!doctype html>. Include hero, specs table, comparison cards, FAQ accordion, and polished CSS. Finish the file with </html>.

Generated tokens: 3,311
Output: 11,953 chars
Total time: 386.430 s
TTFP: 3.843 s
Generated speed: 8.57 tok/s
Draft accepted: 2,094 / 2,436 (86.0%)

Worker 141,904 generated tokens

Create a complete single-file HTML game called Target Pop. Output code only starting with <!doctype html>. Circles appear randomly, the player clicks them for score, timer counts down, and JavaScript handles state. Finish the file with </html>.

Generated tokens: 1,904
Output: 7,306 chars
Total time: 254.617 s
TTFP: 3.858 s
Generated speed: 7.48 tok/s
Draft accepted: 1,218 / 1,370 (88.9%)

Worker 15956 generated tokens

Create a complete single-file HTML Three.js tunnel animation from a CDN. Output code only starting with <!doctype html>. Use rings or boxes moving toward the camera, color cycling, and resize handling. Finish the file with </html>.

Generated tokens: 956
Output: 3,705 chars
Total time: 114.832 s
TTFP: 3.862 s
Generated speed: 8.33 tok/s
Draft accepted: 599 / 712 (84.1%)

Worker 162,508 generated tokens

Create a complete single-file HTML animated explainer page about prompt processing vs token generation. Output code only starting with <!doctype html>. Use CSS animation, diagrams made with divs, and interactive tabs. Finish the file with </html>.

Generated tokens: 2,508
Output: 9,642 chars
Total time: 346.930 s
TTFP: 3.858 s
Generated speed: 7.23 tok/s
Draft accepted: 1,596 / 1,822 (87.6%)

Worker 172,802 generated tokens

Create a complete single-file HTML control panel for a spaceship. Output code only starting with <!doctype html>. Include gauges, buttons, animated warning lights, log messages, and JavaScript state changes. Finish the file with </html>.

Generated tokens: 2,802
Output: 9,792 chars
Total time: 371.417 s
TTFP: 3.853 s
Generated speed: 7.54 tok/s
Draft accepted: 1,781 / 2,040 (87.3%)

Worker 181,779 generated tokens

Create a complete single-file HTML game called Memory Grid. Output code only starting with <!doctype html>. Include cards, shuffle logic, score, timer, matched state, and clean CSS animations. Finish the file with </html>.

Generated tokens: 1,779
Output: 6,626 chars
Total time: 221.494 s
TTFP: 3.849 s
Generated speed: 8.03 tok/s
Draft accepted: 1,161 / 1,236 (93.9%)

Worker 191,515 generated tokens

Create a complete single-file HTML Three.js low-poly landscape from a CDN. Output code only starting with <!doctype html>. Include terrain, floating objects, animated light, and camera drift. Finish the file with </html>.

Generated tokens: 1,515
Output: 4,838 chars
Total time: 201.400 s
TTFP: 3.863 s
Generated speed: 7.52 tok/s
Draft accepted: 946 / 1,138 (83.1%)

Worker 202,634 generated tokens

Create a complete single-file HTML looping title-card video for a model demo. Output code only starting with <!doctype html>. Include animated background, headline transitions, metric callouts, and a replay button. Finish the file with </html>.

Generated tokens: 2,634
Output: 9,629 chars
Total time: 361.823 s
TTFP: 4.028 s
Generated speed: 7.28 tok/s
Draft accepted: 1,664 / 1,940 (85.8%)

Worker 213,473 generated tokens

Create a complete single-file HTML SaaS analytics homepage. Output code only starting with <!doctype html>. Include nav, hero, metric cards, chart mockups using canvas, pricing cards, and responsive CSS. Finish the file with </html>.

Generated tokens: 3,473
Output: 12,151 chars
Total time: 386.951 s
TTFP: 3.833 s
Generated speed: 8.98 tok/s
Draft accepted: 2,239 / 2,470 (90.6%)

Worker 222,505 generated tokens

Create a complete single-file HTML game called Keyboard Pilot. Output code only starting with <!doctype html>. The player dodges meteors using arrow keys, score increases over time, and collisions end the round. Finish the file with </html>.

Generated tokens: 2,505
Output: 8,873 chars
Total time: 335.797 s
TTFP: 3.834 s
Generated speed: 7.46 tok/s
Draft accepted: 1,631 / 1,748 (93.3%)

Worker 231,333 generated tokens

Create a complete single-file HTML Three.js crystal sculpture scene from a CDN. Output code only starting with <!doctype html>. Include reflective-looking geometry, orbiting lights, stars, and animation loop. Finish the file with </html>.

Generated tokens: 1,333
Output: 4,255 chars
Total time: 166.844 s
TTFP: 3.857 s
Generated speed: 7.99 tok/s
Draft accepted: 847 / 970 (87.3%)

Worker 242,291 generated tokens

Create a complete single-file HTML interactive timeline page. Output code only starting with <!doctype html>. Include milestones, animated progress, selectable cards, and JavaScript to switch detail panels. Finish the file with </html>.

Generated tokens: 2,291
Output: 8,261 chars
Total time: 326.640 s
TTFP: 4.040 s
Generated speed: 7.01 tok/s
Draft accepted: 1,444 / 1,692 (85.3%)

Worker 252,877 generated tokens

Create a complete single-file HTML finance-style market dashboard. Output code only starting with <!doctype html>. Include cards, sortable table, animated canvas chart, and fake local demo data. Finish the file with </html>.

Generated tokens: 2,877
Output: 9,775 chars
Total time: 372.446 s
TTFP: 3.835 s
Generated speed: 7.72 tok/s
Draft accepted: 1,850 / 2,054 (90.1%)

Worker 262,143 generated tokens

Create a complete single-file HTML game called Color Reflex. Output code only starting with <!doctype html>. Show colored targets, keyboard shortcuts, reaction-time scoring, timer, and restart logic. Finish the file with </html>.

Generated tokens: 2,143
Output: 7,913 chars
Total time: 302.295 s
TTFP: 3.844 s
Generated speed: 7.09 tok/s
Draft accepted: 1,356 / 1,572 (86.3%)

Worker 271,547 generated tokens

Create a complete single-file HTML Three.js particle globe from a CDN. Output code only starting with <!doctype html>. Include particles on a sphere, rotation, hover text, and responsive resize logic. Finish the file with </html>.

Generated tokens: 1,547
Output: 5,700 chars
Total time: 204.087 s
TTFP: 3.858 s
Generated speed: 7.58 tok/s
Draft accepted: 970 / 1,152 (84.2%)

Worker 282,208 generated tokens

Create a complete single-file HTML animated comic panel sequence. Output code only starting with <!doctype html>. Use CSS keyframes, speech bubbles, scene transitions, and a play/pause button. Finish the file with </html>.

Generated tokens: 2,208
Output: 7,622 chars
Total time: 313.313 s
TTFP: 4.004 s
Generated speed: 7.05 tok/s
Draft accepted: 1,396 / 1,624 (86.0%)

Worker 292,910 generated tokens

Create a complete single-file HTML documentation page for a local benchmark run. Output code only starting with <!doctype html>. Include side nav, code blocks, result cards, collapsible sections, and copy buttons. Finish the file with </html>.

Generated tokens: 2,910
Output: 9,292 chars
Total time: 377.382 s
TTFP: 3.841 s
Generated speed: 7.71 tok/s
Draft accepted: 1,838 / 2,144 (85.7%)

Worker 302,836 generated tokens

Create a complete single-file HTML game called Circuit Clicker. Output code only starting with <!doctype html>. Include click targets, upgrades, score per second, animated circuit board CSS, and JavaScript state. Finish the file with </html>.

Generated tokens: 2,836
Output: 9,641 chars
Total time: 368.461 s
TTFP: 3.839 s
Generated speed: 7.70 tok/s
Draft accepted: 1,835 / 2,004 (91.6%)

Worker 312,651 generated tokens

Create a complete single-file HTML Three.js data-center scene from a CDN. Output code only starting with <!doctype html>. Include glowing server racks, moving light pulses, labels, and camera animation. Finish the file with </html>.

Generated tokens: 2,651
Output: 8,449 chars
Total time: 366.510 s
TTFP: 4.014 s
Generated speed: 7.23 tok/s
Draft accepted: 1,659 / 1,984 (83.6%)

Worker 322,285 generated tokens

Create a complete single-file HTML cinematic loading animation for a 32-slot model demo. Output code only starting with <!doctype html>. Include animated slots, progress bars, status captions, and loop controls. Finish the file with </html>.

Generated tokens: 2,285
Output: 8,342 chars
Total time: 319.893 s
TTFP: 3.829 s
Generated speed: 7.14 tok/s
Draft accepted: 1,456 / 1,656 (87.9%)

Exact server shape

llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --alias main --host 127.0.0.1 --port 18081 --jinja -c 262144 --reasoning off -ngl 999 -fa on -b 4096 -ub 1024 --no-context-shift -dev Vulkan0 --spec-draft-device Vulkan0 -t 16 -tb 16 -ctk f16 -ctv f16 --spec-draft-type-k f16 --spec-draft-type-v f16 --temp 0.6 --min-p 0.0 --top-p 0.95 --top-k 20 --repeat-penalty 1.0 --cache-ram 8192 --parallel 32 --no-mmproj --no-mmap --spec-draft-model mtp-gemma-4-26B-A4B-it.gguf --spec-type draft-mtp --spec-draft-ngl all --spec-draft-n-max 2 --spec-draft-n-min 0 --spec-draft-p-min 0.0 --metrics