Skip to content
noullama

DOCUMENTATION / Using Noullama

Models & downloads

Find decision models on Hugging Face, manage local GGUFs, and follow download progress.

Download from the Hub

  1. Open Hub and search for a decision model. You can also enter organization/repository to open a repository directly.
  2. Select a result on the left to inspect its files on the right.
  3. Choose a model weight GGUF and start the download. Projector files are separate; download one if you intend to use vision.

Discovery uses decision-model tags and architecture metadata. Downloads use hf download at the inspected repository revision.

Authentication

Public models do not require login. For gated models, obtain access on Hugging Face and authenticate. Noullama supports standard Hugging Face token locations and HF_TOKEN. A token entered in the app is held for the session and passed to hf through its environment.

Install the CLI using the official Hugging Face instructions.

Read the memory-fit estimate

Search results estimate memory fit using the repository’s smallest non-projector GGUF. Inspect a repository to see estimates beside individual weight files. Projectors are excluded, and missing metadata displays Fit unavailable.

The estimate uses your current context-size and GPU-layer settings. It is a rough guide, not a guarantee that a model will load.

Use a local model

Open Models → Browse local model to select an existing GGUF, then click Load model. Existing files are used in place. Files containing mmproj are not listed as models; a model with a projector in its folder is marked Vision available.

Set an absolute Model download directory in Models to choose where new downloads go. This does not move existing files. The library scans both that directory and the original app-data models directory, and rescans every few seconds for files added or removed outside Noullama.

Progress, cancellation and retry

The progress bar fills from both ends: network bytes from the left and file reconstruction from the right. Xet compression and deduplication can make the values differ; the ends meet when the file is done.

Choose Cancel download to stop. Partial files are kept for an immediate retry to resume, then deleted when you quit Noullama. While you are elsewhere in the app, the bottom-right Downloads card shows progress and can cancel the download or open its repository.

Repository artwork

Author logos are fetched from public Hugging Face profiles when first displayed and cached for the session. Authors without an image show their initials. No author artwork is bundled with the app.