Open-Weight Models Explained: Weights, Licenses and What Open Means

An open-weight model is one whose trained parameters you can download and run yourself. That is different from fully open source, where training data and training code are also published. Open-weight models are what make local LLMs possible, but each one comes with a license that decides what you may do with it.

What are weights?

Training a language model adjusts billions of numbers, called weights or parameters. After training, those numbers are saved to files. The files plus a small config describing the architecture are everything an inference engine needs to produce text. When you hear "7B model", it means about seven billion of these numbers.

Open weights vs open source

With a traditional open-source program you get source code and the right to modify and redistribute it. With an LLM there are more parts: the training data, the training code, the weights and the license. Many popular models publish only the weights, sometimes with a technical report. You can run, test and often fine-tune them, but you cannot fully reproduce the training. Some projects do release data and code too, and those are closer to open source in the full sense. When reading a model card, check which of these parts are really available.

Common license types

Licenses vary a lot, so read the actual text for any model you plan to ship in a product. These are the general patterns you will meet:

• Permissive licenses such as Apache 2.0 or MIT. You can use the model commercially with few conditions, usually keeping a notice.
• Community or custom licenses from the model's creator. They often allow commercial use but add rules, for example attribution, restrictions on certain uses, or conditions for very large companies.
• Research or non-commercial licenses. Fine for learning and experiments, not for a paid product.
• Acceptable-use policies that list prohibited uses and apply even when the license is otherwise generous.

Post 88 gives a checklist for reviewing a license before commercial use. This post is general information, not legal advice.

Where models come from

The main hub is Hugging Face, which hosts model repositories with a model card, files, and license information. Ollama has its own library of ready-to-run models you fetch with ollama pull. LM Studio has a built-in search that lists models from Hugging Face. Always prefer uploads from the original model creator or from well-known conversion publishers, and check the license field on the model page.

Base, instruct and chat models

A base model is trained only to continue text. An instruct or chat model is further trained to follow instructions and hold a conversation. For almost every local use you want the instruct version. If you load a base model and ask it a question, it may ramble or continue your question instead of answering it.

Formats you will see

• Safetensors: the standard format for full-precision weights on Hugging Face, used by transformers and vLLM.
• GGUF: a single-file format used by llama.cpp, Ollama and LM Studio, often quantized to be smaller.
• Other quantized formats such as GPTQ and AWQ, mainly used with GPU servers.

The next posts explain GGUF and quantization in depth.

Inspecting a model before you download

Once a model is in Ollama, you can read its details and its license text from the command line:

ollama show llama3.2
ollama show llama3.2 --license

The first command prints architecture, parameter count, context length and quantization. The second prints the license that came with the model. Make this a habit: it takes ten seconds and tells you what you are actually running.

Common mistakes

• Assuming "free to download" means "free for any commercial use".
• Downloading a random re-upload instead of the creator's official repository.
• Loading a base model and expecting chat behaviour.
• Forgetting that fine-tuned derivatives inherit the license of the original model.