Models and Responses Help

Models are managed by Geist inside the Pitchblend window. Pitchblend packages the target-specific inference runtime, but it does not bundle or silently download model weights.

Chat says no model is selected

For a local model:

Open Models > Local Models.
Download a local artifact.
Choose Use for that artifact.
Confirm its row is marked Active, then return to Chat.

For an online model, open Settings > Models and Providers and choose the
provider/model catalog entry. The Models provider tab has no Use control,
and selection does not prove credentials or account access. Supply the
provider's required environment credential before starting Pitchblend.

A completed download does not automatically make a model active.

A local model will not download

Confirm that the computer is online and has enough free space for the model plus temporary download work.
Keep Pitchblend open until the Models page reports completion.
Retry from Models after a network interruption.
Check whether the model provider requires you to accept terms or supply a token.

The current curated Windows/Linux choice is a Qwen3 4B Q4_K_M GGUF file of about 2.5 GB. The curated macOS choice is a Meta Llama 3.1 8B Instruct MLX model and is gated by its provider, so its download may require an eligible Hugging Face token.

Do not put provider tokens in chat messages or support screenshots.

Importing a GGUF model

On a GGUF-capable Windows or Linux build, the Models page can import a local .gguf file. Geist validates the extension and file header, then copies the artifact into its managed model store. Importing does not convert other formats or guarantee that every community model will fit your hardware. Keep your own backup of the original file and its source information.

The first response is slow

Local models load lazily when they are first used. The initial response after selecting or switching a model can therefore take longer. On Windows and Linux, the runtime can use Vulkan when compatible drivers are available and fall back to CPU when they are not. CPU fallback is expected to be slower; it is not proof that startup failed.

Wait for one generation to finish or select Stop before sending another prompt. Closing other memory-intensive applications can help when the selected model nearly fills available memory.

Chat stays busy or a response fails

Select Stop if a generation is still active.
Send a short plain-text prompt with no file reference.
If that works, add the file or longer context back separately.
If an online provider was active, switch to an active downloaded local model to isolate the provider from the workspace.
For an online model, verify its provider configuration and network access without exposing the credential.

A failed model response does not automatically mean the whole Geist workspace stopped. Use Startup and Workspace Help only when the wrapper displays a workspace startup or stopped message.

An online provider entry does not work

Provider entries in the model catalog still require the provider's own credentials and network service. The provider, not Pitchblend's local runtime, processes those prompts. Confirm that its credential was available before Pitchblend started, then check account entitlement, network access, and provider service status. The normal interface has no API-key field. Switch to an active downloaded model to continue local inference while diagnosing the online provider.

Responses are too short or low quality

State the task, desired format, important facts, and constraints explicitly.
Start a fresh chat when unrelated earlier context is influencing the answer.
Use a file reference only when the document contains information needed for the request.
Check the active model and its context limit on the Models page.
Review generation values in Settings.

If responses became unexpectedly short after selecting Reset to Defaults, inspect Max Tokens. The current reset action sets it to 16, which can produce very short answers; choose a value appropriate for the active model and task.

For model controls and settings details, see Models, Chat and Composer, and Settings.