Back

Images, 3D and language models on a single RTX 5090

3 November 2026 · Damien Mazeas

Image, 3D and language models on a dedicated RTX 5090 workstation, reachable from a phone. No bill per call, and the prompts never leave the machine.

The studio’s 3D tab: a generated model of a technician in a high-visibility vest and AR glasses, on the viewer’s white background
The 3D tab: a textured model of about 120,000 faces in around 40 s.
The studio's image tab: a prompt box, aspect ratio and step buttons on the left, a generated picture of a car on a wet street on the right
The image tab. This one took 7.6 s; the button promises about 3.

What runs on it

GPU
RTX 5090, 32 GB
Images
FLUX.3 [klein], INT8
Upscale
DAT2 / Nomos2, 4x-UltraSharp, to 4096 px
3D
BiRefNet, then TRELLIS.2
Coding assistant
Gemma 4 E4B, Muse Glimmer, Qwen3.6 35B-A3B
Orchestration
ComfyUI, a Python service, an OpenAI-compatible API
Access
Cloudflare Worker and Tunnel
Hardware link
C# (.NET 8), NVML, Corsair iCUE, Windows LampArray
The language models serve as a local coding assistant through the OpenAI-compatible API. They don’t fit together, so each is loaded on demand.
The dashboard: 1.3 of 31.8 GB of video memory in use, the GPU idle at 30.8 °C and 16 W, and the three language models with their size and request counts
The dashboard at rest. 413 requests so far, 271 of them to the 28B model.

No open port

The machine only makes outgoing connections. No port is open on its network.

  1. Phonebrowser
  2. Workerpassword
  3. Tunnelcloudflared
  4. Python:8085
  5. ComfyUI:8188
Each sampling step comes back over a WebSocket to the progress bar. 3D files (8 to 13 MB) are kept in the browser's cache for a day.

The RAM as a busy light

A small C# program drives the case LEDs (Corsair iCUE, Windows LampArray) and reads the GPU through NVML. While a job runs, the RAM flashes.