> ## Documentation Index
> Fetch the complete documentation index at: https://forge-64364c0e.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Compress

> Quantize GGUF models across 7 compression levels with real-time size and quality preview

# Compress (03)

Quantize GGUF models to smaller sizes using llama-quantize. Preview estimated size, quality, and speed before running.

<img src="https://mintcdn.com/forge-64364c0e/ndaDrHy7hzGW88bh/images/light/compress-light.png?fit=max&auto=format&n=ndaDrHy7hzGW88bh&q=85&s=f3009b837133523be544816c7cb14681" alt="Compress Module" width="3440" height="1406" data-path="images/light/compress-light.png" />

<Note>
  Requires a loaded **GGUF** model and **llama.cpp tools** installed (see [Settings](/modules/settings)).
</Note>

## Quantization Levels

| Level    | Type     | Bits/Weight | Quality | Use Case                                    |
| -------- | -------- | :---------: | :-----: | ------------------------------------------- |
| EXTREME  | Q2\_K    |     2.63    |  \~60%  | Smallest possible, significant quality loss |
| TINY     | Q3\_K\_S |     3.50    |  \~68%  | Very small, noticeable quality loss         |
| SMALL    | Q3\_K\_M |     3.91    |  \~72%  | Small with better quality                   |
| COMPACT  | Q4\_K\_M |     4.85    |  \~80%  | Good balance for most use cases             |
| BALANCED | Q5\_K\_M |     5.69    |  \~87%  | Recommended general purpose                 |
| HIGH     | Q6\_K    |     6.56    |  \~93%  | Near-original quality                       |
| ULTRA    | Q8\_0    |     8.50    |  \~98%  | Minimal quality loss                        |

## Presets

<CardGroup cols={3}>
  <Card title="MOBILE" icon="mobile">
    **Q3\_K\_M** — Edge devices, phones, low-RAM systems
  </Card>

  <Card title="BALANCED" icon="scale-balanced">
    **Q5\_K\_M** — General purpose desktops and laptops
  </Card>

  <Card title="QUALITY" icon="star">
    **Q8\_0** — Production servers, quality-critical applications
  </Card>
</CardGroup>

## Size & Quality Preview

Before quantizing, a preview shows:

* **Before/After** file sizes with reduction percentage
* **Component breakdown**: attention, MLP, embeddings, output head, norms
* **Quality estimate** bar
* **Speed improvement** relative to original

<Warning>
  Requantizing an already-quantized model produces worse results than quantizing from a high-precision source (F16/F32).
</Warning>

## Workflow

<Steps>
  <Step title="Load a GGUF model">
    Use the [Load module](/modules/load) to import a GGUF file
  </Step>

  <Step title="Select target level">
    Click a level button or preset card
  </Step>

  <Step title="Review preview">
    Check estimated size, quality, and speed
  </Step>

  <Step title="Quantize">
    Click **QUANTIZE MODEL**, choose output path, monitor progress
  </Step>
</Steps>

<Tip>
  The original model is never modified — output is always a new GGUF file.
</Tip>