Local Models

Local Models let the main 3sparks Chat app download supported GGUF models and run inference on device with the bundled llama.cpp runtime. This is different from LM Studio or Ollama, which run on another computer or server.

Download and load a model

1. Open Local Models in 3sparks Chat.
2. Browse Available models or use the filter field.
3. Download a supported model.
4. Open Downloaded after the download finishes.
5. Tap Load.
6. Start chatting with the loaded local model.

Catalog behavior

The available model list refreshes from Hugging Face and focuses on public GGUF files the app knows how to prompt safely. The app filters out oversized files, split shards, unsupported quantizations, gated/private repos, and files that are not usable as chat models.

If Hugging Face is unavailable, the app falls back to a starter catalog.

Download a model that is not in the catalog

You are not limited to the curated list. Use Add Model by ID or URL in Available to download any public Hugging Face GGUF model.

1. Open Local Models and go to Available.
2. Tap Add Model by ID or URL.
3. Paste a model reference and tap Look Up Model.
4. Review the repository details and pick a file.
5. Tap Download.

The sheet also links to Browse GGUF Models on Hugging Face if you want to find a repository first and copy its ID.

What you can paste

• A model ID, such as bartowski/Llama-3.2-1B-Instruct-GGUF
• A model page URL, such as https://huggingface.co/Qwen/Qwen3-0.6B-GGUF
• A direct file URL ending in .gguf, which preselects that file

Short hf.co links and tree, blob, and resolve paths all work. Dataset URLs, model listing pages, and non-Hugging Face addresses are rejected.

Choosing a file

After a successful lookup, the app lists every single-file GGUF in the repository with its size and quantization, ordered with the most commonly useful quantizations first. Lower-bit quantizations such as Q4 are smaller and faster. Higher-bit ones such as Q8 and F16 are larger but more accurate.

Split multi-part shards, projector files, and importance-matrix files are not offered, because the app runs single-file GGUF models. Vision projector files are still paired automatically when the model supports images.

Untested models

Custom downloads are not part of the tested catalog, and the sheet says so. The app cannot verify that a model outside the catalog will download, load, or run correctly on your device. The catalog's size limit does not apply either, so a custom file can be larger than anything the curated list offers. The app warns when a model may be too large for the device, but the warning is an estimate.

If a custom model does not work, delete it and try a different quantization or a different repository.

Lookup errors

• Couldn't find it — check the model ID or URL for typos.
• Gated or private — the repository requires Hugging Face access approval and cannot be downloaded directly. Look for an open GGUF upload of the same model.
• No single-file GGUF this app can run — the repository has no usable file. Look for a repository with GGUF in its name, such as one from bartowski or QuantFactory.

Storage and downloads

Downloaded models are stored in the app sandbox and excluded from iCloud backup. Model downloads use iOS background transfers so downloads can continue while the phone is locked or the app is suspended.

Limitations

• Only one Local Model is loaded at a time.
• Larger models need more memory and may be slow on older devices.
• Local Models are main-app only; the keyboard extension cannot use the main app's loaded on-device runtime.

Use LM Studio, Ollama, or another reachable server if the keyboard needs local-model access.

Troubleshooting

• Start with smaller models if you are unsure what your device can run.
• Keep enough free storage for the model file and download overhead.
• If a model fails to load, delete it, refresh the catalog, and try a different model.
• If generation fails, unload and reload the model before retrying.