I Made Nemotron Write 7 Graphics Demos: 59 to 3.6 FPS, No GPU (Nemotron 3 Ultra, OpenRouter)

Unofficial test. I gave NVIDIA's Nemotron 3 Ultra (free tier on OpenRouter) seven prompts for graphics demos and ran every result in Chromium on a machine with no graphics card. Three worked on the first try. Four needed fixes, hints, or a second look, and I show every failure, every retry and every frame rate. The code is not published as a repo, see below. Measured results (all with nvidia/nemotron-3-ultra-550b-a55b:free, thinking off, cost $0, 29 Oct 2026): Plasma: first try, 60 lines in 39.6 s, 59.4 FPS average in the browser Star field: first try, 122 lines in 59.5 s, 50.2 FPS Boids: first try drew wrong wrap-around lines, fixed on try 2 (I pasted the problem back), 6.2 FPS Mandelbrot: first try, 157 lines, 29.2 FPS Terrain: two tries showed only a purple sky, a third try with my hint worked but ran at 3.6 FPS Galaxy: thinking on timed out at 600 s with no code, two tries gave a white screen (a pixel format bug), a fourth try with my hint worked at 8.4 FPS Ray marching: first try gave a black screen (shader error), second try after pasting the error ran at 33.7 FPS 7 demos took 14 attempts. 3 worked first try. Frame rates are software rendering on a shared machine, so treat them as relative, not as a benchmark Plasma prompt, thinking off vs on: 39.6 s vs 61.8 s. Nemotron 3.5 Lightning on the same prompt: 300 s gateway timeout with thinking off, 592.9 s with thinking on (free tier, one run each) Code: there is no public repo. The prompts are shown on screen, and the demo HTML files can be shared on request. Chapters 0:00 The plan: seven prompts, no GPU, honest results 0:49 Setup: key, prompts, rules 1:16 The Nemotron models on OpenRouter 2:23 Why frame rates here are low 2:46 Demo 1: plasma 3:45 Demo 2: 3D star field 4:49 Demo 3: boids (a fix and a failure) 6:06 Demo 4: Mandelbrot 7:00 Demo 5: terrain fly-over 8:10 Demo 6: spiral galaxy (it went badly) 9:25 Demo 7: ray marching 10:31 Speed: tokens per second 10:54 Thinking on vs off, Ultra vs Lightning 11:40 Final scoreboard 13:00 Honest notes 13:41 References 13:59 Outro How this video was made The voice is synthetic text to speech (Microsoft Edge TTS, en-US-AndrewNeural) reading a script I wrote. The seven prompts were written before recording. The fix prompts after a failure were also written by me, and some contain hints. That is why those demos are marked as fixed. The browser runs are real screen recordings. The terminal views are replays of the logged model responses, shown at their original timing and labelled REPLAY on screen, so the text is readable. The code views show the real saved files in VS Code. Mouse and keyboard input was scripted. Waiting is sped up and always labelled SPED UP. Browser frame rates were measured inside the page with no graphics card, so they are lower than on a normal PC. Animated cards were rendered in Python (Pillow) and edited with FFmpeg. The OpenRouter key was never shown. Free models are rate limited and change often. This is a small sample, not a ranking. References Nemotron on OpenRouter: https://openrouter.ai/nvidia OpenRouter API docs: https://openrouter.ai/docs/quickstart Streaming: https://openrouter.ai/docs/api/refere... Models list (checked 29 Oct 2026): https://openrouter.ai/models NVIDIA Nemotron on Hugging Face: https://huggingface.co/nvidia edge-tts: https://github.com/rany2/edge-tts Credits and licenses Music: original background track synthesized in Python (NumPy) for this video, no samples. Fonts: Poppins (SIL Open Font License), DejaVu Sans Mono (free license). Nemotron is a trademark of NVIDIA. This channel is not affiliated with or sponsored by NVIDIA or OpenRouter. No NVIDIA logo is used. #Nemotron #NVIDIA #OpenRouter #AIcoding #JavaScript #WebGL #creativecoding