Learn to use GPU Flow, step by step.

1.1What is an inference API?

"Inference" is simply the moment an already-trained AI model answers a question. An "API" is the door through which your software talks to other software. Put together: your app sends a request with your text, our server passes it to the model running on a GPU and returns the answer, usually in under a second and, if you want, word by word (streaming).

The good part: because we speak the same "language" as OpenAI and Anthropic, you don't rewrite your code. You change the address it points to and that's it.

The round-trip of one request:

And "tokens"? The model doesn't read letters or whole words but chunks called tokens (a short word is usually 1 token; a long one, 2-3). You count input tokens (what you send) and output tokens (what it answers). That's how price is computed:

Three things worth understanding about API consumption:

• 1Input and output are billed separately. Output usually costs more per token: generating text is more expensive than reading it.
• 2Each call is independent. The model remembers nothing between requests. To hold a conversation you resend the history on every call, and that history also counts as input tokens, so long conversations consume more.
• 3Streaming doesn't change the price. You pay the same tokens; you just see the answer appear sooner, word by word.

1.2Get started in 3 minutes

1.3Your first call

Using the OpenAI Python SDK. Only two lines change: base_url and api_key.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.gpuflow.ai/v1",
    api_key="gpf-...",
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Hola, mundo"}],
    stream=True,
)

for chunk in response:
    print(chunk.choices[0].delta.content or "", end="")

curl https://api.gpuflow.ai/v1/chat/completions \
  -H "Authorization: Bearer gpf-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-3.5-9b",
    "messages": [{"role": "user", "content": "Hola, mundo"}],
    "stream": false
  }'

1.4Main endpoints

Base URL: https://api.gpuflow.ai/v1 · authenticate with Authorization: Bearer gpf-…

1.5Model catalogue

Models served today in Spain. Prices in EUR per million tokens, VAT included, split into input (your prompt) and output (the response). The ★ marks featured models.

• ANot sure which? Start with a ★ featured model (DeepSeek V4 Pro): a good balance of quality and price for most tasks.
• BHigh volume or simple replies: pick a small model (Qwen3.5 9B, Gemma 4 E2B): cheaper and almost instant.
• CReasoning or complex code: use the larger ones (DeepSeek V4 Pro, GLM 5.2 FP8). They cost more per token but get it right more often.

Need another model? If it fits the GPU, we'll deploy it; write to [email protected].

1.6Credit & billing

Prepaid model: top up whenever you like and it's consumed by real usage (tokens). Each top-up comes with a Spanish VAT invoice.

The amount you pay is shown on top; the credit you receive includes the bonus. Higher volume or B2B usage? Enterprise tiers exist; just ask.

2.1What is a Sandbox?

You connect to it over SSH, a secure connection, with your own editor or terminal, and install whatever you need. It's ideal for developing, running notebooks and wiring your agents to inference, with no infrastructure to set up.

The key to not overspending is understanding its three states and what's billed in each:

2.2Why on our cluster

A GPU Flow Sandbox isn't just any remote machine: it runs on the same sovereign cluster as the rest of the platform, and that's where its advantages come from.

More on certifications and data residency on the web security page.

2.3Key concepts

2.4Create a Sandbox, step by step

2.5Connecting over SSH

SSH is the standard way to log into a remote computer securely. Instead of a username and password it uses a key pair: two files that only work together. Understanding this pair is the only "new" thing you need:

Step 1: generate your key (if you don't have one). The file without .pub is your private key: never share it. You only paste the .pub into the dashboard.

# 1 · create the keys folder
New-Item -ItemType Directory -Force -Path "$HOME\.ssh" | Out-Null
# 2 · generate your key (Enter for no passphrase)
ssh-keygen -t ed25519 -C "gpuflow"
# 3 · copy & paste this to get your key and view it again later
Get-Content "$HOME\.ssh\id_ed25519.pub"

# 1 · create the keys folder
mkdir -p ~/.ssh && chmod 700 ~/.ssh
# 2 · generate your key (Enter for no passphrase)
ssh-keygen -t ed25519 -C "gpuflow"
# 3 · copy & paste this to get your key and view it again later
cat ~/.ssh/id_ed25519.pub

Paste the line starting with ssh-ed25519 into the wizard's public-key field.

Step 2: connect. You go through our public bastion, no VPN needed. The bastion checks your key's fingerprint and routes you to your Sandbox.

# connect via the bastion (no VPN) · replace <tu-entorno> with your Sandbox name
ssh -J ssh.gpuflow.ai developer@devpod-<tu-entorno>

2.6Persistent volume lifecycle

This is the most surprising and most useful part: deleting the Sandbox does not delete your data. The volume survives and you reuse it whenever you want.

2.7How it's billed

Three simple rules, and you'll always see the estimated cost before confirming:

3.1What is a dedicated GPU?

It is a GPU reserved exclusively for your project, as opposed to a Sandbox where the accelerator is assigned on demand. It is the right fit when you have a continuous workload (long training, production inference) and want stable, predictable performance.

3.2Configurations

We offer the full B200 (180 GB) and MIG fractions (mutually isolated slices of a B200) for smaller workloads. Exact configurations and prices are on the pricing page; product detail on the dedicated GPU page.

3.3How to request it

In the dedicated GPU configurator you pick the GPU model; the rest of the options are fixed for now. On submit, we ask for your details and contact you to reserve and set up your node. Once active, you connect over SSH just like a Sandbox (same bastion, same public key).

3.4How it's billed

As reserved capacity, pricing is tailored to the configuration and reservation period — we confirm it during contact. Reference rates (full B200 and MIG fractions, hourly or reserved) are on the pricing page.

3.1General concepts

3.2Sandboxes & infrastructure