scaled to zero
- request
- find a free gpu
- load the weights
- warm up
- first token
quarry, reserved
- request
- already loaded on your gpus
- first token
the model waits, not you.
Open models on reserved GPUs, at one fixed price per million tokens.
already warm, every call.
Each deployment keeps its model loaded on its own reserved GPUs, so the first request answers like the hundredth.
- support-botwarm
- search-rerankwarm
- draft-writerwarm
- new deployment
support-bot
warm- model
- open-model-70b
- gpus
- 4 reserved
- cold start
- 0 s
- rate
- fixed per 1m tokens
promptWhere is my order?
It shipped this morning; the tracking link is in your inbox.
| time | tokens in | tokens out | status |
|---|---|---|---|
| 09:41:07 | 212 | 58 | 200 |
| 09:41:03 | 188 | 41 | 200 |
| 09:40:52 | 304 | 96 | 200 |
the rate holds at 3 a.m.
Traffic moves. The price per million tokens doesn’t.
cold start
0s
The weights stay loaded between calls, so no request waits for a GPU.
price per million tokens
fixed
One figure for every token, whatever the traffic did.
one request, plain https.
To your own warm deployment. No client library to install.
curl https://api.example.com/v1/generate \
-H "Authorization: Bearer $QUARRY_API_KEY" \
-d '{ "deployment": "support-bot",
"prompt": "Where is my order?" }'import os, requests
r = requests.post(
"https://api.example.com/v1/generate",
headers={"Authorization": f"Bearer {os.environ['QUARRY_API_KEY']}"},
json={"deployment": "support-bot", "prompt": "Where is my order?"},
)
print(r.json()["output"])const r = await fetch("https://api.example.com/v1/generate", {
method: "POST",
headers: { Authorization: `Bearer ${process.env.QUARRY_API_KEY}` },
body: JSON.stringify({ deployment: "support-bot", prompt: "Where is my order?" }),
});
console.log((await r.json()).output);response
{ "output": "It shipped this morning; the tracking
link is in your inbox.",
"usage": { "input_tokens": 212, "output_tokens": 58 } }- status
- 200
- deployment
- warm
- cold start
- 0 s