The machine
What this box actually offers a local model.
There is no placeholder reading anywhere on this page. A VRAM gauge that shows 17.2 / 24 GB when nothing has been measured is worse than no gauge, because it is indistinguishable from a working one. probe/machine_probe.py reports what it finds and stays silent about what it does not.
Graphics
NOT REPORTED
nvidia-smi is not on this machine. That is not the same as having no GPU, and it is certainly not the same as having 0 GB of VRAM. Nothing is reported.
Local model
ANSWERING
1 local model endpoint(s) answered on loopback.
Models named: gpt-oss:20b, llama3.1:latest
| Runtime | Port | Port open | Answered | Models | Note |
|---|---|---|---|---|---|
| NVIDIA NIM (OpenAI-compatible) | 8000 | closed | no | — | |
| NVIDIA NIM embeddings | 8080 | closed | no | — | |
| Ollama | 11434 | open | yes | gpt-oss:20b, llama3.1:latest | |
| llama.cpp server | 8081 | closed | no | — | |
| vLLM | 8001 | closed | no | — | |
| Text generation inference | 3000 | closed | no | — |
Probed 2026-08-15T13:11:32+00:00 on macOS-26.4.1-arm64-arm-64bit-Mach-O · 16 logical CPUs. A port being open is not the same as a model answering, and both are reported separately for exactly that reason — a build server squatting on port 3000 will open the port and answer nothing a model would say.
Report what the hardware says. Report the silence too.
Run it on your own box
Three commands.
# 1. what does this machine actually have
python3 probe/machine_probe.py
# 2. rebuild the catalog from whatever is in the staging folder
node tools/ingest.mjs && node tools/extract-claims.mjs
# 3. serve the studio locally, where the local model is reachable
python3 tools/serve.py # then open http://127.0.0.1:8787/The console only works from a local origin. A page served from the public internet cannot reach a model on your loopback address — your browser blocks it, correctly, and no amount of code on this end changes that. That restriction is a feature: it means this deployed copy could not quietly use your local model even if it wanted to. Serve the studio from 127.0.0.1 and the console connects.
Your own hardware, in the browser
Read from your browser, not from the build. Nothing is sent anywhere.
Reading…