Measured artifact comparison
Read straight off the three artifacts running on this host. Nothing here is a score, a ranking, or a judgement — it is what each build contains and does. Open a row's disclosure to see how the number was taken.
How this is served
One container per model
Every clone is a node:22-alpine image that copies that model's
artifact into /app and runs node server.js. No clone's
source was patched, renamed, or rewritten; the only knobs are environment
variables and container security options (read_only,
cap_drop: ALL, non-root, no published port). Caddy terminates TLS
with Let's Encrypt and is the only thing that listens on 80/443.
Why three hostnames
Each clone asks for its own assets and API by absolute path
(/app.js, /api/puzzle). Serving them under one host with
a path prefix would break those requests, so each clone keeps its own
subdomain. This hub reaches a clone's health endpoint only, through
/probe/<clone>/api/health, which is why the status above works
without changing any clone.
The models
Two are different models; the third is the same architecture as the second at a smaller quantisation, so the second and third columns isolate quantisation while the first changes the model family. The IQ3_XXS build was served locally by llama.cpp on four RTX 5060 Ti cards with NVMe-backed offload and a 240k context.