Open-weight models

Open-weight models are AI models whose trained weights are published for download, so you can run and fine-tune them on your own hardware. The license, not the download, decides what you may do with them.

What it means

A model is open-weight when its publisher releases the trained parameters for download, usually on the Hugging Face Hub. You can load them, run them on your own GPUs and change them. A closed model is reachable only through its vendor's API.

Open weight is not open source

The weights are one part of a model. The training data and the training code are often not released with them. Licenses vary too: some models use permissive licenses such as Apache 2.0 or MIT, and others carry custom terms that limit commercial use or ask you to accept conditions before downloading. Read the license before you build on a model.

Why it matters for fine-tuning

Fine-tuning changes weights, so you need to hold them. With an open-weight model you choose the method, from LoRA to full fine-tuning, keep your data on your own machines, and keep the tuned weights. A closed model can be tuned only as far as its vendor's API allows.

In Halo

Halo trains Hugging Face models as they are. Any AutoModelForCausalLM trains with no extra code: point model_name_or_path at it and launch. The result loads with from_pretrained. The supported models page lists the families with added parallelism support.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL