DiggTop
Menu
⌕

Set your run

No downloads, no scripts. Copy one prompt, send it to your local AI, and watch each step land here live — machine, model, then the benchmark. Confirm and it is committed to the board.

Your ticket: run-CZC3-HPD8 — keep it safe; it is your claim key.

1. Fill the prompt

local AI prompt Download .txt
You are submitting a local LLM benchmark to diggtop.com using curl only — do NOT download any files.
Your ticket is run-CZC3-HPD8. The page shows CONNECTED the moment your first POST lands, then keep going step by step. Complete these three steps IN ORDER, posting each result to https://www.diggtop.com/api/v1/submit/step as JSON, then reply with ONLY the JSON you posted.

STEP 1 (machine config) — detect the GPU/CPU and VRAM (nvidia-smi, system_profiler, lscpu), then POST:
{"ticket":"run-CZC3-HPD8","step":"config","data":{"hardware":{"name":"RTX 3090","vendor":"nvidia","vram_gb":24},"driver_os":"Linux"}}

STEP 2 (model) — report the loaded model and engine, then POST:
{"ticket":"run-CZC3-HPD8","step":"model","data":{"model":{"name":"Qwen3.8-27B","family":"Qwen","params_b":28},"quant":"Q4_K_M","backend":"llama.cpp","backend_ver":"b4105","context_len":4096}}

STEP 3 (benchmark) — measure single-stream generation tok/s against the local server, then POST:
- warm up with one generation request
- run 3 timed generations, divide generated tokens by seconds, take the median
- read peak VRAM with nvidia-smi --query-gpu=memory.used (or free on macOS)
{"ticket":"run-CZC3-HPD8","step":"bench","data":{"tps_avg":42.1,"prompt_tps":120.0,"peak_vram_gb":21.5}}

Reply with ONLY the JSON of each step as you complete it.

If you cannot make HTTP requests, just print the three JSON objects (one per line) and tell the user to paste them into the page.

2. Watch your run land

Waiting to connect to your local AI…

Usually under a minute. If it stays this way, your agent probably can’t reach diggtop.com — a proxy or Cloudflare challenge is the usual culprit.

  1. Fill the rig — machine config
  2. Fill the model — model + quant
  3. Run the bench — benchmark

Waiting for the first step… (the page polls every 3s)

Stuck on “Waiting to connect…”?

Your agent must be able to POST to https://www.diggtop.com/api/v1/submit/step. If it runs behind a proxy, keep diggtop.com out of it (e.g. NO_PROXY=diggtop.com,www.diggtop.com) — some proxy egress IPs get challenged by Cloudflare. Then have your agent retry the first step.

No agent? Run the three curls yourself
curl -X POST https://www.diggtop.com/api/v1/submit/step -H "Content-Type: application/json" -d '{"ticket":"run-CZC3-HPD8","step":"config","data":{"hardware":{"name":"RTX 3090","vendor":"nvidia","vram_gb":24},"driver_os":"Linux"}}'

curl -X POST https://www.diggtop.com/api/v1/submit/step -H "Content-Type: application/json" -d '{"ticket":"run-CZC3-HPD8","step":"model","data":{"model":{"name":"Qwen3.8-27B","family":"Qwen","params_b":28},"quant":"Q4_K_M","backend":"llama.cpp","backend_ver":"b4105","context_len":4096}}'

curl -X POST https://www.diggtop.com/api/v1/submit/step -H "Content-Type: application/json" -d '{"ticket":"run-CZC3-HPD8","step":"bench","data":{"tps_avg":42.1,"prompt_tps":120.0,"peak_vram_gb":21.5}}'
AI only printed the JSON (no HTTP)? Paste it here
Reads the ticket from the pasted objects — any step already received lights up instantly.

3. Claim & share

After confirming, open your run to claim it (add a handle, a note and a link) and grab a shareable card:

Submissions are pending review — methodology (engine, version, quant, context) is checked before they reach the board.

Fresh community runs

↑