How to Load or Convert Models in LlamaChat

Updated

LlamaChat does not ship with any model files. To use it, you must obtain a model yourself and then bring it into the app in one of two forms: a raw published PyTorch checkpoint, which LlamaChat can import and convert, or a model you have already converted to the .ggml format. Either way, you are responsible for following the terms set by whoever provides the model.

What you need before starting

  • LlamaChat v1.2.0 installed on a Mac running macOS 13 or later. The build supports both Intel processors and Apple Silicon.
  • A model source. LlamaChat is built to chat with LLaMA, Alpaca, and GPT4All models, with Vicuna listed as coming soon. The app's own description notes that Alpaca is Stanford's 7B-parameter LLaMA model fine-tuned on instruction-following demonstrations, which is why it behaves more like a chatbot than base LLaMA.
  • The provider's license or terms. Since no model files are distributed with the app, check the conditions attached to whichever model you download before you integrate it.

Option 1: Import a raw PyTorch checkpoint

LlamaChat can import raw published PyTorch model checkpoints directly. This is the path to take if you have downloaded the original weights rather than a pre-converted file.

  1. Download the model checkpoint from its publisher and keep the files together in the folder where they were delivered.
  2. Open LlamaChat and start an import of that checkpoint.
  3. Let the app handle the conversion. Because the checkpoint is being converted rather than merely opened, expect this step to take noticeably longer than loading a ready-made file, and give it time to finish before interacting with the model.

The expected result is a model that appears in LlamaChat and is ready to chat with, without you having to run a separate conversion tool.

Option 2: Import a pre-converted .ggml file

If you already have a .ggml model file, or you converted one outside the app, LlamaChat can import it as-is. This skips the conversion work inside the app and is the faster route when a .ggml build of your chosen model is available.

  1. Confirm the file is in .ggml format.
  2. Import it into LlamaChat.
  3. Select the model and begin chatting.

Choosing between the two

Situation Better option
You only have the original published weights Raw PyTorch checkpoint import
You already have or can easily get a .ggml file Pre-converted .ggml import
You want to avoid a long conversion step Pre-converted .ggml import
No .ggml build exists for your model Raw PyTorch checkpoint import

How to verify it worked

After either route, the model should be selectable in LlamaChat and should respond to a test message. If it does not appear, recheck that the files are complete and that the format matches one of the two supported types. If a raw checkpoint import stalls, the conversion is likely still running rather than failed.

Common sticking points

  • Expecting bundled models. None are included, proprietary or otherwise, so an empty model list on first launch is normal.
  • Wrong format. Only raw PyTorch checkpoints and .ggml files are supported import types; other formats need converting first.
  • Licensing. Obtaining and integrating model files in line with each provider's terms is your responsibility, not something the app handles for you.
  • macOS version. The current release requires macOS 13, so older systems will not run it.

LlamaChat itself is free and fully open-source, powered by llama.cpp and llama.swift, and its source is available on GitHub if you want to check how conversion is handled or contribute changes.