> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fastrun.si/llms.txt
> Use this file to discover all available pages before exploring further.

# Introduction

> Fast inference for open models through an OpenAI-compatible API.

fastrun serves open models on dedicated GPUs and answers through the same API as OpenAI. If your code already uses
an OpenAI SDK, switching takes three changes: the base URL, the API key and the model name.

| | |
| - | - |
| Base URL | `https://api.fastrun.si/v1` |
| Authentication | `Authorization: Bearer <your fastrun key>` |
| Endpoints | `/chat/completions`, `/completions`, `/models` |
| Model | `google/gemma-4-31b-it` |

> [card] [Quickstart](https://docs.fastrun.si/quickstart.md)
>
> Make your first request in a couple of minutes.

> [card] [Models](https://docs.fastrun.si/models.md)
>
> What is served, with its limits and features.

> [card] [Streaming](https://docs.fastrun.si/guides/streaming.md)
>
> Show tokens as they are generated.

> [card] [API reference](https://docs.fastrun.si/api-reference/introduction.md)
>
> Every endpoint and parameter.

## What you get

* **Speed.** Gemma 4 31B runs on NVIDIA B200 GPUs with speculative decoding. In our own measurement, one request at a
  time without network time, the median time to first token was 24 ms and the median output speed was 280 tokens per
  second. Latency rises under heavy load.
* **OpenAI compatibility.** Chat completions, streaming, tool calling, JSON output and reasoning work with the official
  OpenAI SDKs and the tools built on them.
* **Pay per token.** You pay for input and output tokens only, with a lower price for input that is served from the
  prompt cache. See [Pricing](https://docs.fastrun.si/pricing.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.