> ## Documentation Index
> Fetch the complete documentation index at: https://docs.os3compute.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Training

> Train a cheaper model of your own on your traffic, when you are ready.

When you are ready, train a smaller open model on your own requests to do your tasks for less. It is yours alone.

## Start a run

**1. Opt in**

An owner or admin turns on **Processing** and **Allow training** in **Settings → Training data**. Nothing trains before.

**2. Choose**

**Train new version** in the portal: pick a model below and a dataset of your requests.

**3. Start**

Check the estimated charge, then **Start training**.

## Models you can train

| Model | Training |
| - | - |
| Qwen3.8-27B | GPU time; per token soon |
| Qwen3.6-35B-A3B | GPU time |
| Qwen3.5-9B | GPU time |
| Qwen3.5-4B | GPU time |
| gemma-4-31B-it | GPU time |
| gemma-4-12B-it | GPU time |
| Qwen3.8-Flash-Next-FP8 | GPU time; per token soon |

Each has a context of 262,144 tokens.

## How training is billed

Per-token pricing starts soon for the models below. Until then runs are billed by GPU time, US\$3.99 per GPU-hour.

US\$ per 1M training tokens, launch prices for runs started before 2027-02-15:

| Model | Examples up to 65,536 tokens | Longer examples |
| - | - | - |
| Qwen3.8-27B | 2.05 | 3.73 |
| Qwen3.8-Flash-Next-FP8 | 6.33 | 4.00 |

## Serving a trained model

It runs on OS3's GPUs, billed per token. Its releases are in **Models**.

US\$ per 1M tokens, from 2026-11-18:

| Trained from | Input | Cached input | Output |
| - | - | - | - |
| Qwen3.8-27B, Qwen3.8-Flash-Next-FP8 | 0.18 | 0.0063 | 0.72 |

Models trained from other bases: [ask us](https://docs.os3compute.com/support.md).


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