MODEL PLAYGROUND

The live tool runs on the machine itself

Browser-side API calls from this documentation page are blocked by browser private-network policy (public site → tailnet IP), so the interactive playground is served from the machine itself over the tailnet HTTPS endpoint — same origin as the API, no CORS, no workarounds:

https://frame-serv-2.tail8056c7.ts.net/playground/

tailnet device required (any device on the dxverm tailnet)

• Chat-style: type a prompt, hit run — the same prompt goes to each selected model one at a time. Each run appears as a log entry with its status.
• Output links: when a model finishes you get an open output link — full text, first-token latency, total time, token count, t/s, and the complete transcript. Compare outputs side by side at your own pace.
• Tool-use mode: optional second round — the request carries a get_weather tool; structured tool calls get executed (simulated weather) and fed back, so you see the full two-step agent loop.
• Saved locally: run history lives in your browser's localStorage — past runs re-appear with their output links on every visit.

Timing note: the first token includes the model swap (llama-swap unloads the previous model and loads the next — 67–113 GB swaps take minutes). The t/s on the output page reflects generation rate after that first token.

1. Paste the endpoint API key into the key field (stored in that browser's localStorage only — never published here; get it from the operator or /srv/llm/secrets/api-key on the machine).
2. Click Load models — the list populates from /v1/models.
3. Optionally enable tool-use test, set max_tokens, then Run sequential benchmark.

Full benchmark methodology: Benchmarks.

Served by the box itself: static page at /srv/llm/playground/ mounted via tailscale serve at /playground/, same origin as the API — no cross-origin requests, no public exposure. Source is a single HTML file; tweak it in place and refresh.