> For the complete documentation index, see [llms.txt](https://docs.wproofreader.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.wproofreader.com/deployment/installation-requirements/hardware-requirements.md).

# Hardware requirements

WProofreader SDK Server can run on:

* A dedicated server
* A virtual machine
* A Docker container
* A cloud provider such as AWS, Azure, or GCP

VMware virtualization is supported.

{% hint style="warning" %}
AI-based language models are not supported under VirtualBox.
{% endhint %}

Resource usage depends on the components you install and the languages you enable.

GEC models, Autocomplete, list of installed languages are selected during installation.

### Disk space

#### Base installation

The base installation uses about **1.2 GB** in the application directory.

It includes:

* AppServerX binary and libraries
* Algorithmic grammar engine and core language resources
* Style guide resources
* Front-end static assets

#### Optional components

| Component                                     | Disk space                                                                                                                                              |
| --------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
| AI GEC model — English                        | \~683 MB                                                                                                                                                |
| AI GEC model — German                         | \~1.3 GB                                                                                                                                                |
| AI GEC model — Spanish                        | \~1.3 GB                                                                                                                                                |
| Autocomplete model — English                  | \~490 MB                                                                                                                                                |
| NER model — per language                      | \~336 MB. See [Configure NER models](https://docs.wproofreader.com/deployment/configuration/spell-check-engine/configure-ner-models.md)                 |
| N-gram model — per language                   | Varies by language. See [Configure n-gram models](https://docs.wproofreader.com/deployment/configuration/spell-check-engine/configure-n-gram-models.md) |
| Hunspell spelling dictionaries — per language | Varies. All languages use \~256 MB total                                                                                                                |

By default, the installer downloads the English GEC model and the English Autocomplete model.

The installer downloads each AI model as an archive of up to about 1 GB and unpacks it in place. Keep about **1 GB** of extra free space while the models are being installed.

#### Full installation

A full installation with all GEC models, English Autocomplete, all NER models, and all n-gram models uses about **6.5 GB** in the application directory.

Hunspell dictionaries are stored separately in the data directory.

#### Additional variable space

| Item                                              | Space                     |
| ------------------------------------------------- | ------------------------- |
| User dictionaries (`UserDictionaries/`)           | Up to 50 KB per user      |
| Organization dictionaries (`CustomDictionaries/`) | Up to 5 MB per dictionary |
| Style guide collections                           | Up to 5 MB per collection |
| AppServer log files (`AppServer/Logs/`)           | Rotated at 10 MB per file |

Web server access logs can also grow quickly. Monitor them separately.

### RAM

#### Recommended configuration

| Configuration                                                                | RAM         |
| ---------------------------------------------------------------------------- | ----------- |
| Default installation: English, English GEC and English Autocomplete, 4 vCPUs | **\~5 GB**  |
| Full installation: all languages, 3 GEC models and English Autocomplete      | **\~14 GB** |
| Spell check and grammar only, no AI                                          | \~2 GB      |

{% hint style="info" %}
By default, only the English GEC model and the English Autocomplete model are installed.
{% endhint %}

{% hint style="warning" %}
With AI models enabled, RAM usage also grows with the number of CPU cores available to the server. The AI engine starts more threads on more cores, and each thread that processes a request needs its own working memory. For example, the default English installation uses about 3.9 GB with 4 vCPUs and about 4.7 GB with 16 vCPUs. Add about **1 GB** to the values above for servers with 8 or more vCPUs.
{% endhint %}

#### Breakdown by component

| Component                                                 | RAM usage     |
| --------------------------------------------------------- | ------------- |
| Spell check engine, per active language                   | \~50 MB       |
| Algorithmic grammar check engine, English                 | \~0.5–1 GB    |
| Algorithmic grammar check engine, all supported languages | \~4 GB        |
| AI engine runtime, shared by all AI models                | \~300 MB      |
| AI GEC model, English                                     | \~1.5 GB      |
| AI GEC model, German or Spanish                           | \~2.4 GB each |
| Autocomplete model, English                               | \~800 MB      |
| N-gram model, per active language                         | \~100 MB      |
| NER model, per active language                            | \~336 MB      |
| Cache, 500,000 misspelling entries                        | \~200 MB      |

Notes:

* The algorithmic grammar check engine for English uses about 0.5 GB with one English dialect and about 1 GB with all four (en\_US, en\_GB, en\_CA, en\_AU).
* The JVM heap limit is **3 GB** by default.
* If you enable more than 15 languages, increase the heap to **4 GB**. See [Configure JVM maximum heap size](https://docs.wproofreader.com/deployment/configuration/grammar-check-engine-setup/configure-jvm-maximum-heap-size.md).
* NER and n-gram models are loaded only for languages selected during installation.
* The English NER model is loaded in the default setup.
* German and Spanish NER models are disabled by default.
* RAM usage grows with concurrent users, active languages, and enabled features.

### CPU

| Item        | Value       |
| ----------- | ----------- |
| Minimum     | 2 vCPUs     |
| Recommended | **4 vCPUs** |

{% hint style="info" %}
**4 vCPUs** are recommended when AI language models are enabled.

**AVX2** or **AVX512** can improve AI model performance.
{% endhint %}

### Cloud instances

| Use case                                             | Suggested instance                               |
| ---------------------------------------------------- | ------------------------------------------------ |
| Light load or evaluation                             | AWS `t3.large` or `t3a.large`: 2 vCPUs, 8 GB RAM |
| Production workload, English GEC                     | AWS `c8a.xlarge`: 4 vCPUs, 8 GB RAM              |
| Production workload, all languages without AI models | AWS `m8a.xlarge`: 4 vCPUs, 16 GB RAM             |
| Production workload, all languages and all AI models | AWS `m8a.2xlarge`: 8 vCPUs, 32 GB RAM            |
| High AI load with batching                           | AWS `g4dn.xlarge` or equivalent, NVIDIA T4 GPU   |

{% hint style="info" %}
AWS `c8a.xlarge` is sufficient for English AI workloads.

AWS `t3.medium` or `t3a.medium` (4 GB RAM) is enough only without AI models.
{% endhint %}

{% hint style="info" %}
For high-throughput AI workloads, NVIDIA T4 Tensor Core GPUs improve batch processing.
{% endhint %}

### Scaling considerations

Actual resource usage depends on:

* Number of concurrent users
* Volume and frequency of checked text
* Enabled languages and their resources
* Percentage of errors in submitted text
* Enabled AI models, NER, and Autocomplete

### Admin-panel setup

The requirements above apply to WProofreader Server in both setups. If you use admin-panel, you also need resources for admin-panel and MySQL. For a comparison of the setups, see [Choose a self-hosted setup](https://docs.wproofreader.com/deployment/installation/choose-a-self-hosted-setup.md).

Admin-panel and MySQL need fewer resources than WProofreader Server.

#### Admin-panel

| Item | Minimum | Recommended |
| ---- | ------- | ----------- |
| CPU  | 1 vCPU  | **2 vCPUs** |
| RAM  | 2 GB    | **4 GB**    |

The values apply to all three admin-panel containers together: the web application, the queue worker and the scheduler. The recommended values are sufficient for production.

Admin-panel keeps uploaded files, such as profile photos, in `storage/app`. When you turn on persistent storage, the Helm chart requests a 5 GB volume.

#### MySQL

| Database                              | CPU     | RAM  |
| ------------------------------------- | ------- | ---- |
| `cloud_service` (WProofreader Server) | 2 vCPUs | 2 GB |
| `admin_panel_db` (admin-panel)        | 1 vCPU  | 1 GB |

You can keep the two databases on two MySQL servers or on one MySQL server. For one MySQL server, use 2 vCPUs and 3 GB RAM.

The size of the `cloud_service` database grows with the request statistics. Monitor the free disk space of the MySQL server.

#### Docker Compose on one host

When you run all components on one host, add the requirements of each component. For example, the default WProofreader Server installation (about 5 GB), admin-panel (2 GB) and MySQL (3 GB) need at least 10 GB RAM.
