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

# Aria model bundles and quantization

> Learn how Aria Compute packages model bundles with quantization levels, SDK versioning, and authenticated delivery via presigned URLs.

An Aria model bundle is a directory that contains everything needed to run a device-native foundation model on your hardware. Each bundle is named by combining the model slug with its quantization level, such as `gemma-4-e2b-it_q4` or `qwen3.5-2b_q8`. This naming convention makes it easy to identify the exact model and precision you are deploying.

## Bundle contents

Every bundle contains at least the following files:

| File          | Purpose                          |
| ------------- | -------------------------------- |
| `weight.bin`  | The quantized model weights      |
| `config.json` | Model configuration and metadata |

Optional tokenizer sidecars such as `tokenizer.json` may also be included, depending on the model architecture.

## Quantization levels

Aria Compute supports three quantization values for model weights:

| Quantization | Description                                                |
| ------------ | ---------------------------------------------------------- |
| `int4`       | 4-bit integer weights, smallest size                       |
| `int8`       | 8-bit integer weights, balanced size and accuracy          |
| `int326`     | INT3.26 mixed precision for language and VLA/policy models |

Choose `int4` when storage is limited, `int8` for balanced quality, and `int326` for advanced language and vision models that require higher precision.

## SDK versioning

When you download a bundle, you can specify an SDK version with the `sdk` parameter (for example, `sdk=v1.0`). This ensures that you receive a bundle that is compatible with the Aria SDK you are using in your application. If you omit the parameter, the server returns the latest compatible bundle.

## Download and registry

Model bundles are hosted on a private S3 registry. Authenticated download requests return a short-lived S3 presigned URL. You can authenticate with either a session JWT or an API key in the `Authorization: Bearer` header.

Read the [Downloading models guide](/guides/download-models) for step-by-step instructions, or see the [`GET /api/models/{slug}/download`](/api-reference/models/download) reference for query parameters and response formats.

## Public mirrors

Aria Compute also maintains public mirrors on Hugging Face (`ariacompute/*`) and ModelScope (`AriaCompute/*`) for users who prefer to download through those platforms.


## Related topics

- [Download Aria model bundles from the registry](/guides/download-models.md)
- [GET /api/models — list available Aria model bundles](/api-reference/models/list.md)
- [GET /api/models/{slug}/download — download an Aria model bundle](/api-reference/models/download.md)
- [aria-engine runtime and FFI](/sdks/engine-runtime.md)
- [Aria Compute Documentation](/index.md)