Models
Open Models — Providers to manage model files on this computer and inspect
models reported by online providers. Select Refresh to reload both local
artifact status and the provider catalog.
Local Models
The Local Models tab opens first. Its Local model files table has four
columns: Artifact, Format, Status, and Actions.
The curated artifact depends on the runtime available for the operating
system:
- supported Windows and Linux builds use Qwen3 4B Q4_K_M (GGUF) through
llama.cpp; and - supported Apple silicon builds use Meta Llama 3.1 8B Instruct (MLX).
Unsupported artifacts are hidden. A gated MLX download can say Requires
accepted Hugging Face access and an HF token. The Models page has no token
entry field, so that access must already be configured for the local runtime.
Download and activate a model
Use the action that matches the current status:
- Download starts the managed download.
- Cancel requests cancellation of a queued or active download.
- Clear removes a partial download.
- Use makes an installed artifact the default local model.
- Active marks the artifact already selected.
- Remove deletes an installed artifact from managed model storage.
Busy rows show downloaded bytes or snapshot-file progress. Errors appear in
the row or in a notice above the table.
The active model cannot be removed. Its tooltip says Select another model
before removing this one. Activate a different installed artifact first.
Selecting Use changes the default runtime to local and records the concrete
artifact. The model itself is loaded lazily when Chat sends its first request,
so that first answer can take longer and may show a model-loading card.
Import a GGUF file
On a GGUF-capable build, select Import GGUF and choose one .gguf file.
Pitchblend validates the extension and GGUF header, then copies the file into
its managed model store. The label changes to Importing… during the
request.
An imported artifact can be activated or removed, but its Download action
is disabled because there is no curated remote source to download it from
again. Keep your own backup of the original model file. There is no model
export action in the app.
Model Providers
The Model Providers tab is a catalog, not a model picker. It shows:
- Active provider, Active model, and Providers summary cards;
- a section and model count for each provider;
- Model, Context, Output, and Capabilities columns; and
- capability badges such as Streaming, Tools, Vision, and
Recommended.
Choose the online provider and model under
Settings, not on this tab. Neither
page provides an API-key or credential field. Selecting an online model does
not by itself grant provider access or prove that its credentials are valid.
Fallback provider data
If live discovery fails, the page says Live model discovery failed, so Geist
is showing fallback model data. The built-in fallback lists GPT-4-family,
Claude 3-family, Grok 2 and 3, Groq Llama 3.3 70B Versatile, and one offline
Meta Llama model.
Treat that list as reference data only. A fallback entry does not guarantee
that the provider still offers the model, that your account can use it, or
that Pitchblend has credentials for it. During a normal live load, the page
shows Loading model providers — Gathering local and online provider
options.