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

# anaconda ai download

## Usage

```bash theme={null}
anaconda ai download <MODEL> [OPTIONS]
```

## Description

Downloads a quantized model file or safetensors collection from the backend. If the file is already downloaded, the command exits immediately unless `--force` is specified. A progress bar displays during the download.

## Download types

### Quantized models

A quantized model is a compressed GGUF file optimized for local inference. To download a quantized model, provide the model name and quantization type as `ModelName/QuantizationType`.

### Safetensors collections

> [!NOTE]
>
> Safetensors collection downloads require the `ai-catalyst` backend.
>
> ***
>
> You cannot launch a server or run inference on safetensors files directly from `anaconda ai`. Use a third-party framework to load and run the model after downloading.

A safetensors collection contains all the files needed to run a model with a third-party inference framework, such as Hugging Face `transformers`. You can download a safetensors collection using the `--safetensors` flag.

A collection typically includes:

File

Purpose

`*.safetensors`

Model weight files (might be sharded across multiple files for large models)

`model.safetensors.index.json`

Index file for sharded weight files

`tokenizer.json`, `tokenizer_config.json`

Tokenizer vocabulary, merge rules, and configuration

`config.json`, `generation_config.json`

Model architecture and default generation parameters

`chat_template.jinja`

Chat formatting template (if applicable)

## Arguments

Argument

Description

`MODEL`

Required Model identifier. Format depends on download type. <ul><li>Quantized model: `ModelName/QuantizationType`</li><li>Safetensors collection: `ModelName` only</li></ul>

## Options

Option

Default

Description

`--safetensors`

`false`

Download the safetensors collection for a model. Requires the `ai-catalyst` backend.

`--force`

`false`

Force re-download if the model is already downloaded

`--output, -o <PATH>`

For quantized downloads: hard-link the model file to this path. For safetensors downloads: directory to download files into (created if it does not exist).

`--backend <BACKEND>`

config value

Select the inference backend

`--at <SITE>`

Target a specific site defined in your configuration

`--json`

`false`

Print output as JSON

`-h, --help`

Show help for this command

## Examples

**Download a model**

```bash theme={null}
anaconda ai download OpenHermes-2.5-Mistral-7B/Q4_K_M
```

**Force re-download of an existing model**

```bash theme={null}
anaconda ai download OpenHermes-2.5-Mistral-7B/Q4_K_M --force
```

**Download and hard-link to a custom path**

```bash theme={null}
anaconda ai download OpenHermes-2.5-Mistral-7B/Q4_K_M -o ./models/openhermes.gguf
```

**Download a safetensors collection**

```bash theme={null}
anaconda ai download --safetensors Qwen3-4B-Thinking-2507
```

**Download a safetensors collection to a specific directory**

```bash theme={null}
anaconda ai download --safetensors Qwen3-4B-Thinking-2507 --output ./my-model
```