pitchblend AI
English
English Spanish

First Launch and Model Setup

Pitchblend starts its bundled private Geist service before it opens the
workspace. Startup can take time while the service and local database
initialize, but Pitchblend does not silently download model weights or run a
setup wizard.

1. Wait for the workspace

The screen shows Loading... while the service reports its loopback address
and becomes ready. Assistive status text can announce Preparing your
workspace… or Reconnecting to your workspace…. Leave the window open.

If you see Geist couldn’t start, Workspace is stopped, or Required UI
couldn’t load, read the detail and choose Retry startup once. If the same
message returns, copy its non-sensitive text and use
Startup and workspace troubleshooting.

2. Open Models

Select Models in the left navigation, then stay on Local Models. The recommended artifact depends on your platform:

  • Windows or Linux x64: Qwen3 4B Q4_K_M (GGUF), about 2.5 GB.
  • Apple silicon macOS: Meta Llama 3.1 8B Instruct (MLX). This gated model requires accepting its Hugging Face terms and supplying a Hugging Face token before Pitchblend starts.

The installer contains the model runtime, not the model weights. Choose Download, keep Pitchblend open, and wait for the status to finish. The download needs an internet connection and enough free disk space.

3. Make it active

Choose Use beside the installed model. Its row changing to Active is
the confirmation that a concrete managed artifact is selected. The top bar
only summarizes saved catalog settings and can name a model that is not
installed.

The model is loaded lazily when the first chat needs it, so the first response can take longer. Windows and Linux attempt Vulkan acceleration and fall back to CPU automatically; macOS Apple silicon uses MLX. You do not need to choose an accelerator.

4. Try a small prompt

Open Chat, type:

In two sentences, explain what a local AI model is.

Press Enter or choose Send. If text streams into the Geist response, the basic local path is working.

What to postpone

Keep the first test simple. Do not add files, voice, an online provider, web search, a workflow, account sign-in, or Memory Sync until plain local chat works. Adding one feature at a time makes any problem much easier to identify.