CotrievalAI

Docs

Book a machine and use it

Each booking makes an NVIDIA machine for you alone, loads the model, gives you a way in, and deletes the machine when your time ends.

Book a machine

  1. Sign in at the portal with Google, or with your email: we send a six-digit code, and there is no password. Signing in the first time creates your account.
  2. Choose the Confidential GPU or GLM-5.3-Flash, the machine, and 1 to 6 hours. The price is shown before you pay.
  3. Pay through Stripe Checkout. The purchase is sold through Link, which sends your receipt.
  4. Wait while it is prepared. The booking page shows each step as the machine reports it: 5 to 10 minutes for the Confidential GPU, 10 to 15 for GLM-5.3-Flash, once a zone has a free GPU. If a zone has none, the booking tries the next one, which can add several minutes.
  5. Use it. When it is ready, your booked hours start, and the page shows how to reach it. We also email you.
  6. Add hours from the booking page while it runs, up to 6 hours in all. Each addition is its own payment, and the end time moves as soon as Stripe confirms it. We email you 10 minutes before the end.
  7. End it whenever you like from the booking page, or let the time run out. The machine and its disk are deleted.

GLM-5.3-Flash

An uncensored build of GLM-5.3-Flash (glm-5.3-flash-uncensored) served by vLLM 0.31 on four GPUs, with a 1M-token context, tool calling and reasoning output. The API is OpenAI-compatible, so OpenCode, the OpenAI SDKs and curl work with it.

The booking page has a chat box, Try it here, that talks to your machine straight from your browser. For your own tools, it shows three values once the machine is ready: the base URL (your machine's own address, for example https://34.0.0.1/v1), your API key (ct-…, for this booking only), and the model name. The address is an IP with its own Let's Encrypt certificate, so HTTPS works without a domain.

curl

curl https://YOUR-MACHINE/v1/chat/completions \
  -H "Authorization: Bearer $COTRIEVAL_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3-flash-uncensored","messages":[{"role":"user","content":"Hello"}]}'

Python

import os
from openai import OpenAI

client = OpenAI(base_url="https://YOUR-MACHINE/v1", api_key=os.environ["COTRIEVAL_KEY"])
reply = client.chat.completions.create(
    model="glm-5.3-flash-uncensored",
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)

OpenCode

{
  "provider": {
    "cotrieval": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "https://YOUR-MACHINE/v1", "apiKey": "{env:COTRIEVAL_KEY}" },
      "models": { "glm-5.3-flash-uncensored": {} }
    }
  }
}

The machine serves up to 16 requests at once. Speed for one user: 226–231 tokens a second on 4× H100, about 145 on 4× RTX PRO 6000 (measured Oct 14).

Confidential GPU

One RTX PRO 6000 Blackwell (96 GB) in a Google Confidential VM, running ComfyUI 0.39 in your browser. Press Open ComfyUI on your booking page: the link carries a token for this booking only and opens the machine's own address.

Start with a ready-made workflow

Open ComfyUI's Workflows panel (press W). Six workflows are there, each ComfyUI's own template for the models on the machine:

  1. Picture from text (Qwen-Image 2.1, with its prompt enhancer)
  2. Edit a picture (Qwen-Image 2.1)
  3. Remove a background (Qwen-Image 2.1)
  4. Video with sound from text (MiniMax H3)
  5. Video with sound from a picture (MiniMax H3)
  6. Video with sound from reference pictures (MiniMax H3)

Write your prompt, add your pictures where a workflow asks for them, and press Run. The Templates list shows only the templates whose models are on the machine, including segmentation with SAM 3 and BiRefNet background removal. ComfyUI opens on the picture workflow. Measured on a booked machine on Oct 17: the first 1024 px picture took 20 seconds, model loading included, and a 5-second 864×480 video with sound took about 2.5 minutes.

What is on it

Once it is ready the machine has no outbound internet, so ComfyUI cannot download models or install nodes. Uploads, results and saved workflows live only in its encrypted memory and are gone when the booking ends: download what you want to keep.

Time and refunds

What happensWhat you pay
Your time runs outThe hours you booked
You end it earlyThe time you used; the rest is refunded automatically
Google reclaims the Spot machineThe time you used; the rest is refunded
No zone has a free machine, or it never becomes readyNothing; everything is refunded
You cancel before it is readyNothing; everything is refunded

Refunds go back to the card you paid with; the bank usually shows them within 5 to 10 days. Every payment and refund is listed under Billing in the portal, with its receipt.

Limits