Mirid docs

Begin without guesswork.

This is the short route from installer to conversation. Mirid exposes deeper controls when you need them; none are a toll gate on your way in.

First launch

Install Mirid, open it, and leave the window running while the local engine arrives. Mirid downloads roughly 3.1 GB, verifies the archive, and extracts a runtime of about 9 GB.

  1. Keep about 14 GB of disk space free during setup.
  2. Keep Mirid open until the progress screen completes.
  3. If the connection stops, close and reopen Mirid. The download resumes rather than starting again.
  4. When onboarding appears, choose hosted APIs, a local model, or both.
No GPU? Mirid should still open. Choose API-only use for the lightest path, or run smaller GGUF models on the CPU if you are comfortable waiting longer.

Use hosted models

Hosted providers are the simplest way to begin on an ordinary laptop. Mirid does not sell an account of its own; you add a key issued by the provider you choose.

  1. Open Settings → Providers.
  2. Choose a provider and follow its link to create or view an API key.
  3. Paste the key into Mirid, then refresh that provider’s model list.
  4. Select a model in the model library and return to the conversation.

NanoGPT is highlighted for broad text, image and audio access. OpenRouter is useful when you want one account across many model developers. Direct provider connections remain available when you prefer them.

Subscription access is not the same as an API key. A consumer chat subscription from a provider may not include API credit. Mirid shows links and guidance, but billing remains between you and that provider.

Run a local model

Mirid runs GGUF text models through llama.cpp. You can select a file already on your machine or use the Hugging Face downloader.

Choose a file you already have

  1. Open Settings → Local models.
  2. Select your .gguf file.
  3. Review the estimated memory use at the context lengths Mirid offers.
  4. Choose a context that fits your available memory, then load the model.

Download from Hugging Face

Paste a repository name such as publisher/model-name. If you know the exact quantisation, append its filename or choose it from the files Mirid discovers. A Q4 quantisation is often a sensible first attempt on modest hardware.

Model weights are separate from Mirid’s 3.1 GB runtime and can range from a few gigabytes to tens of gigabytes. The model’s repository should state its licence and intended use.

Create or import a character

Open the character library to start from a blank character, build one through Mirid’s guided creator, or import a compatible character card.

  • Guided creator: describe the idea in ordinary language and refine identity, voice, boundaries and opening scene through conversation.
  • Character card: import supported PNG or JSON cards used by the wider roleplay ecosystem.
  • Dataset import: map a collection of character records when its columns contain recognisable names, descriptions, greetings or example dialogue.

Read imported material before using it. Character repositories are community content; their quality, permissions and safety vary.

Add voice and images

Hear a reply

Choose a voice in Settings → Audio, then press the play control beneath a message. Mirid shows that synthesis has started while the audio engine works. Kokoro is included in the core runtime; larger voice engines remain optional downloads.

Bring images into chat

Use the image attachment control to select one or more files, or drag images directly into the message area. The active model must support vision for the images to affect its reply.

Generate an image

Configure a supported hosted image model or a local Stable Diffusion model. Local image generation benefits greatly from a compatible GPU and requires additional model files.

Connect other applications

Mirid can expose an OpenAI-compatible local endpoint so software such as SillyTavern can use the model currently available through Mirid.

  1. Open Settings → Server and enable the compatibility server.
  2. Copy the endpoint address and access key shown by Mirid.
  3. In the other application, choose an OpenAI-compatible or custom endpoint.
  4. Paste the address and key, then select the Mirid model it discovers.

Only expose the server beyond your own machine if you understand the network and authentication settings. A local endpoint can control models and incur provider charges using your configured credentials.

Troubleshooting

The first launch appears stuck

Large downloads can sit at one percentage while a chunk completes. Leave Mirid open. If there is no network activity for several minutes, relaunch it to resume. Verification begins after the download and can also take time on slower storage.

Mirid opens an Edge “page not found” screen

That indicates a development or incomplete build rather than the release installer. Remove that copy and install the current release from the download page. The production app embeds its interface and does not need a web server.

A local model will not load

Choose a smaller quantisation or context length. The memory estimate includes the model and context cache, and Windows still needs room for Mirid and other applications.

A hosted model fails

Confirm the provider key, available credit and exact model selection. If model rotation is enabled, make sure at least one model is included in the rotation—or turn rotation off to use the single selected model.