# Inference quickstart

GraphN serves open models behind an OpenAI-compatible API. If you already use an OpenAI SDK, change two things: the base URL and the API key. New accounts start with $1 of inference credit.

## 1. Create an API key

In the app, open **API keys** at the bottom of the sidebar and create a key. Keys start with `gn_` and belong to one workspace; requests made with a key are billed to that workspace's organization.

Export it so the examples below can read it:

```bash
export GRAPHN_API_KEY=gn_...
```

## 2. Send a request

**Base URL:** `https://model.graphn.ai/v1`

### curl

```bash
curl https://model.graphn.ai/v1/chat/completions \
  -H "Authorization: Bearer $GRAPHN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-80b",
    "messages": [{"role": "user", "content": "Hello"}]
  }'
```

### Python

Install the SDK with `pip install openai`, then:

```python
import os

from openai import OpenAI

client = OpenAI(base_url="https://model.graphn.ai/v1", api_key=os.environ["GRAPHN_API_KEY"])

response = client.chat.completions.create(
    model="qwen3-80b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)
```

### TypeScript

Install the SDK with `npm install openai`, then:

```typescript
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://model.graphn.ai/v1",
  apiKey: process.env.GRAPHN_API_KEY,
});

const response = await client.chat.completions.create({
  model: "qwen3-80b",
  messages: [{ role: "user", content: "Hello" }],
});
console.log(response.choices[0].message.content);
```

### CLI

```bash
brew tap voltagepark/graphn && brew install graphn
graphn login
graphn model chat qwen3-80b -m "Hello"
```

`graphn model chat` without `-m` starts an interactive session. See [Developing with agents](/docs/developing-with-agents) for other install options.

## 3. Pick a model

The `model` field takes the model's callable name, such as `qwen3-80b` or `deepseek-v4.1-flash`. [Models](/docs/models) lists every built-in model with its context window, and each links to a page with pricing. To list the models your key can call:

```bash
curl https://model.graphn.ai/v1/models \
  -H "Authorization: Bearer $GRAPHN_API_KEY"
```

Models you import or deploy yourself are called the same way, as `imported:<id>` or `custom:<id>`. The **Playground** in the app runs any of them and shows the matching request in its **Code** view.

## Streaming

Set `stream` to receive the reply as server-sent events, one chunk at a time:

```python
stream = client.chat.completions.create(
    model="qwen3-80b",
    messages=[{"role": "user", "content": "Write a haiku about GPUs."}],
    stream=True,
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)
```

## Images

Vision models take OpenAI-style content parts. An image can be an `https` URL or a `data:` URL:

```python
response = client.chat.completions.create(
    model="qwen3-vl",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "What is in this image?"},
                {"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}},
            ],
        }
    ],
)
```

## Parameters beyond the OpenAI SDK

Some models take parameters the OpenAI SDK does not define, such as `top_k` or a thinking switch. In Python, pass them through `extra_body`:

```python
response = client.chat.completions.create(
    model="qwen3.8-27b",
    messages=[{"role": "user", "content": "Hello"}],
    extra_body={"top_k": 20, "chat_template_kwargs": {"enable_thinking": False}},
)
```

Reasoning tokens count toward `max_tokens`, so a thinking model given a small budget can stop before it answers. Turn thinking off, or raise the budget.

## Errors

| Status | Meaning                                                                            |
| ------ | ---------------------------------------------------------------------------------- |
| `401`  | The key is missing, revoked, or malformed.                                         |
| `402`  | The organization is out of credit. Add credit under **Billing**.                   |
| `404`  | No model by that name is available to this workspace.                              |
| `429`  | Too many requests. Retry with backoff.                                             |
| `503`  | A model you deployed is scaled to zero and is waking up. Retry after a short wait. |

## Next steps

- Compare models side by side in the Playground's **Compare** view.
- [Deploy your own weights](/docs/tutorial-custom-model-sdk) or [attach LoRA adapters](/docs/lora-adapters).
- Build multi-step pipelines on top of these models with [workflows](/docs/getting-started).